├── .flake8 ├── .github └── workflows │ └── sphinx.yml ├── .gitignore ├── .pylintrc ├── LICENSE ├── README.md ├── docs ├── Makefile ├── _static │ ├── favicon.ico │ ├── flowchart.png │ ├── icon-alt.png │ ├── icon.png │ ├── logo-alt.png │ └── logo.png ├── _templates │ └── base.html ├── api.rst ├── conf.py ├── index.rst ├── make.bat └── requirements.txt ├── examples ├── README.rst ├── classification.py ├── regression.py ├── trust_classification.py └── trust_regression.py ├── poetry.lock ├── pyproject.toml ├── requirements.txt └── trustee ├── __init__.py ├── _version.py ├── enums ├── __init__.py └── feature_type.py ├── main.py ├── report ├── __init__.py ├── plot.py └── trust.py └── utils ├── __init__.py ├── const.py ├── dataset.py ├── log.py ├── persist.py ├── plot.py ├── rootpath.py └── tree.py /.flake8: -------------------------------------------------------------------------------- 1 | [flake8] 2 | max-line-length = 120 -------------------------------------------------------------------------------- /.github/workflows/sphinx.yml: -------------------------------------------------------------------------------- 1 | name: Sphinx build 2 | 3 | on: 4 | push: 5 | branches: 6 | - master 7 | 8 | jobs: 9 | build: 10 | runs-on: ubuntu-latest 11 | steps: 12 | - uses: actions/checkout@v3 13 | - name: Build HTML 14 | uses: ammaraskar/sphinx-action@master 15 | - name: Upload artifacts 16 | uses: actions/upload-artifact@v3 17 | with: 18 | name: html-docs 19 | path: docs/build/html/ 20 | - name: Deploy 21 | uses: peaceiris/actions-gh-pages@v3 22 | with: 23 | deploy_key: ${{ secrets.ACTIONS_DEPLOY_KEY }} 24 | external_repository: TrusteeML/trusteeml.github.io 25 | publish_dir: docs/build/html 26 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | *.joblib 3 | 4 | __pycache__/ 5 | *.py[cod] 6 | *$py.class 7 | 8 | # C extensions 9 | *.so 10 | 11 | # Distribution / packaging 12 | .Python 13 | auto_examples/ 14 | build/ 15 | develop-eggs/ 16 | dist/ 17 | downloads/ 18 | eggs/ 19 | .eggs/ 20 | lib/ 21 | lib64/ 22 | parts/ 23 | sdist/ 24 | var/ 25 | wheels/ 26 | pip-wheel-metadata/ 27 | share/python-wheels/ 28 | *.egg-info/ 29 | .installed.cfg 30 | *.egg 31 | MANIFEST 32 | 33 | # PyInstaller 34 | # Usually these files are written by a python script from a template 35 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 36 | *.manifest 37 | *.spec 38 | 39 | # Installer logs 40 | pip-log.txt 41 | pip-delete-this-directory.txt 42 | 43 | # Unit test / coverage reports 44 | htmlcov/ 45 | .tox/ 46 | .nox/ 47 | .coverage 48 | .coverage.* 49 | .cache 50 | nosetests.xml 51 | coverage.xml 52 | *.cover 53 | *.py,cover 54 | .hypothesis/ 55 | .pytest_cache/ 56 | 57 | # Translations 58 | *.mo 59 | *.pot 60 | 61 | # Django stuff: 62 | *.log 63 | local_settings.py 64 | db.sqlite3 65 | db.sqlite3-journal 66 | 67 | # Flask stuff: 68 | instance/ 69 | .webassets-cache 70 | 71 | # Scrapy stuff: 72 | .scrapy 73 | 74 | # Sphinx documentation 75 | docs/_build/ 76 | 77 | # PyBuilder 78 | target/ 79 | 80 | # Jupyter Notebook 81 | .ipynb_checkpoints 82 | 83 | # IPython 84 | profile_default/ 85 | ipython_config.py 86 | 87 | # pyenv 88 | .python-version 89 | 90 | # pipenv 91 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 92 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 93 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 94 | # install all needed dependencies. 95 | #Pipfile.lock 96 | 97 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow 98 | __pypackages__/ 99 | 100 | # Celery stuff 101 | celerybeat-schedule 102 | celerybeat.pid 103 | 104 | # SageMath parsed files 105 | *.sage.py 106 | 107 | # Environments 108 | .env 109 | .venv 110 | env/ 111 | venv/ 112 | ENV/ 113 | env.bak/ 114 | venv.bak/ 115 | 116 | # Spyder project settings 117 | .spyderproject 118 | .spyproject 119 | 120 | # Rope project settings 121 | .ropeproject 122 | 123 | # mkdocs documentation 124 | /site 125 | 126 | # mypy 127 | .mypy_cache/ 128 | .dmypy.json 129 | dmypy.json 130 | 131 | # Pyre type checker 132 | .pyre/ 133 | 134 | # Custom files 135 | *.csv 136 | .DS_Store 137 | *.pcap 138 | *.zip 139 | *.md5 140 | dataset/ 141 | *.pdf 142 | dt_* 143 | -------------------------------------------------------------------------------- /.pylintrc: -------------------------------------------------------------------------------- 1 | good-names=i,j,k,ex,Run,_,pk,x,y,X,X_train,X_test 2 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | Trustee 2 | 3 | [![Downloads](https://static.pepy.tech/personalized-badge/trustee?period=total&units=international_system&left_color=grey&right_color=blue&left_text=downloads)](https://pepy.tech/project/trustee) 4 | 5 | 6 | This package implements the `trustee` framework to extract decision tree explanation from black-box ML models. 7 | For more information, please visit the [documentation website](https://trusteeml.github.io). 8 | 9 | Standard AI/ML development pipeline extended by Trustee. 10 | Trustee 11 | 12 | Getting Started 13 | --------------- 14 | 15 | This section contains basic information and instructions to get started with Trustee. 16 | 17 | ### Python Version 18 | 19 | Trustee supports `Python >=3.7`. 20 | 21 | ### Install Trustee 22 | 23 | Use the following command to install Trustee: 24 | 25 | ``` 26 | $ pip install trustee 27 | ``` 28 | 29 | ### Sample Code 30 | 31 | ``` 32 | from sklearn import datasets 33 | from sklearn.ensemble import RandomForestClassifier 34 | from sklearn.model_selection import train_test_split 35 | from sklearn.metrics import classification_report 36 | 37 | from trustee import ClassificationTrustee 38 | 39 | X, y = datasets.load_iris(return_X_y=True) 40 | X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30) 41 | 42 | clf = RandomForestClassifier(n_estimators=100) 43 | clf.fit(X_train, y_train) 44 | y_pred = clf.predict(X_test) 45 | 46 | trustee = ClassificationTrustee(expert=clf) 47 | trustee.fit(X_train, y_train, num_iter=50, num_stability_iter=10, samples_size=0.3, verbose=True) 48 | dt, pruned_dt, agreement, reward = trustee.explain() 49 | dt_y_pred = dt.predict(X_test) 50 | 51 | print("Model explanation global fidelity report:") 52 | print(classification_report(y_pred, dt_y_pred)) 53 | print("Model explanation score report:") 54 | print(classification_report(y_test, dt_y_pred)) 55 | ``` 56 | 57 | ### Usage Examples 58 | 59 | For simple usage examples of Trustee and TrustReport, please check the `examples/` directory. 60 | 61 | ### Other Use Cases 62 | 63 | For other examples and use cases of how Trustee can used to scrutinize ML models, listed in the table below, please check our [Use Cases repository](https://github.com/TrusteeML/emperor). 64 | 65 | | Use Case | Description | 66 | | ------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------- | 67 | | heartbleed\_case/ | Trustee application to a Random Forest Classifier for an Intrustion Detection System, trained with CIC-IDS-2017 dataset pre-computed features. | 68 | | kitsune\_case/ | Trustee application to Kitsune model for anomaly detection in network traffic, trained with features extracted from Kitsune\'s Mirai attack trace. | 69 | | iot\_case/ | Trustee application to Random Forest Classifier to distguish IoT devices, trained with features extracted from the pcaps from the UNSW IoT Dataset. | 70 | | moon\_star\_case/ | Trustee application to Neural Network Moon and Stars Shortcut learning toy example. | 71 | | nprint\_ids\_case/ | Trustee application to the nPrintML AutoGluon Tabular Predictor for an Intrustion Detection System, also trained using pcaps from the CIC-IDS-2017 dataset. | 72 | | nprint\_os\_case/ | Trustee application to the nPrintML AutoGluon Tabular Predictor for OS Fingerprinting, also trained using with pcaps from the CIC-IDS-2017 dataset. | 73 | | pensieve\_case/ | Trustee application to the Pensieve RL model for adaptive bit-rate prediction, and comparison to related work Metis. | 74 | | vpn\_case/ | Trustee application the 1D-CNN trained to detect VPN traffic trained with the ISCX VPN-nonVPN dataset. | 75 | 76 | ### Supported AI/ML Libraries 77 | 78 | | Library | Supported | 79 | | ------------ | ------------------ | 80 | | scikit-learn | :white_check_mark: | 81 | | Keras | :white_check_mark: | 82 | | Tensorflow | :white_check_mark: | 83 | | PyTorch | :white_check_mark: | 84 | | AutoGluon | :white_check_mark: | 85 | 86 | ## Citing us 87 | 88 | ``` 89 | @inproceedings{Jacobs2022, 90 | title = {AI/ML and Network Security: The Emperor has no Clothes}, 91 | author = {A. S. Jacobs and R. Beltiukov and W. Willinger and R. A. Ferreira and A. Gupta and L. Z. Granville}, 92 | year = 2022, 93 | booktitle = {Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security}, 94 | location = {Los Angeles, CA, USA}, 95 | publisher = {Association for Computing Machinery}, 96 | address = {New York, NY, USA}, 97 | series = {CCS '22} 98 | } 99 | ``` 100 | -------------------------------------------------------------------------------- /docs/Makefile: -------------------------------------------------------------------------------- 1 | # Minimal makefile for Sphinx documentation 2 | # 3 | 4 | # You can set these variables from the command line, and also 5 | # from the environment for the first two. 6 | SPHINXOPTS ?= 7 | SPHINXBUILD ?= sphinx-build 8 | SOURCEDIR = ./ 9 | BUILDDIR = build 10 | 11 | # Put it first so that "make" without argument is like "make help". 12 | help: 13 | @$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O) 14 | 15 | .PHONY: help Makefile 16 | 17 | # Catch-all target: route all unknown targets to Sphinx using the new 18 | # "make mode" option. $(O) is meant as a shortcut for $(SPHINXOPTS). 19 | %: Makefile 20 | @$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O) -v 21 | -------------------------------------------------------------------------------- /docs/_static/favicon.ico: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TrusteeML/trustee/71561fbcc96af8596fbd23879c3edcf0639e985b/docs/_static/favicon.ico -------------------------------------------------------------------------------- /docs/_static/flowchart.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TrusteeML/trustee/71561fbcc96af8596fbd23879c3edcf0639e985b/docs/_static/flowchart.png -------------------------------------------------------------------------------- /docs/_static/icon-alt.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TrusteeML/trustee/71561fbcc96af8596fbd23879c3edcf0639e985b/docs/_static/icon-alt.png -------------------------------------------------------------------------------- /docs/_static/icon.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TrusteeML/trustee/71561fbcc96af8596fbd23879c3edcf0639e985b/docs/_static/icon.png -------------------------------------------------------------------------------- /docs/_static/logo-alt.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TrusteeML/trustee/71561fbcc96af8596fbd23879c3edcf0639e985b/docs/_static/logo-alt.png -------------------------------------------------------------------------------- /docs/_static/logo.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TrusteeML/trustee/71561fbcc96af8596fbd23879c3edcf0639e985b/docs/_static/logo.png -------------------------------------------------------------------------------- /docs/_templates/base.html: -------------------------------------------------------------------------------- 1 | {% extends "!base.html" %} 2 | 3 | {% block scripts %} 4 | {{ super() }} 5 | 6 | 7 | 14 | {% endblock %} -------------------------------------------------------------------------------- /docs/api.rst: -------------------------------------------------------------------------------- 1 | API 2 | === 3 | 4 | This part of the documentation covers all the interfaces of Trustee. For 5 | parts where Trustee depends on external libraries, we document the most 6 | important right here and provide links to the canonical documentation. 7 | 8 | .. automodule:: trustee.main 9 | :members: 10 | :undoc-members: 11 | :show-inheritance: 12 | :inherited-members: 13 | :special-members: __init__ 14 | 15 | 16 | .. automodule:: trustee.report.trust 17 | :members: 18 | :undoc-members: 19 | :show-inheritance: 20 | :inherited-members: 21 | :special-members: __init__ 22 | -------------------------------------------------------------------------------- /docs/conf.py: -------------------------------------------------------------------------------- 1 | # Configuration file for the Sphinx documentation builder. 2 | # 3 | # For the full list of built-in configuration values, see the documentation: 4 | # https://www.sphinx-doc.org/en/master/usage/configuration.html 5 | 6 | import os 7 | import sys 8 | from datetime import date 9 | 10 | sys.path.insert(0, os.path.abspath("..")) 11 | 12 | import trustee 13 | 14 | # -- Project information ----------------------------------------------------- 15 | # https://www.sphinx-doc.org/en/master/usage/configuration.html#project-information 16 | 17 | project = "Trustee" 18 | copyright = f"{date.today().year}, Arthur Selle Jacobs" 19 | author = "Arthur Selle Jacobs" 20 | # The short X.Y version. 21 | version = trustee.__version__ 22 | # The full version, including alpha/beta/rc tags. 23 | release = version 24 | 25 | 26 | # -- General configuration --------------------------------------------------- 27 | # https://www.sphinx-doc.org/en/master/usage/configuration.html#general-configuration 28 | 29 | extensions = [ 30 | "sphinx.ext.autodoc", 31 | "sphinx.ext.intersphinx", 32 | "sphinx.ext.coverage", 33 | "sphinx.ext.napoleon", 34 | "sphinxemoji.sphinxemoji", 35 | "sphinx_gallery.gen_gallery", 36 | ] 37 | 38 | 39 | templates_path = ["_templates"] 40 | 41 | # -- Options for HTML output ------------------------------------------------- 42 | # https://www.sphinx-doc.org/en/master/usage/configuration.html#options-for-html-output 43 | 44 | 45 | html_theme = "furo" 46 | html_favicon = "_static/favicon.ico" 47 | html_static_path = ["_static"] 48 | html_theme_options = { 49 | "light_logo": "icon.png", 50 | "dark_logo": "icon-alt.png", 51 | "footer_icons": [ 52 | { 53 | "name": "GitHub", 54 | "url": "https://github.com/TrusteeML/trustee", 55 | "html": """ 56 | 57 | 58 | 59 | """, 60 | "class": "", 61 | }, 62 | ], 63 | } 64 | 65 | sphinx_gallery_conf = { 66 | "examples_dirs": "../examples", # path to your example scripts 67 | "gallery_dirs": "auto_examples", # path to where to save gallery generated output 68 | } 69 | -------------------------------------------------------------------------------- /docs/index.rst: -------------------------------------------------------------------------------- 1 | .. image:: _static/logo.png 2 | :width: 400px 3 | :align: center 4 | :alt: Trustee 5 | :class: only-light 6 | 7 | .. image:: _static/logo-alt.png 8 | :width: 400px 9 | :align: center 10 | :alt: Trustee 11 | :class: only-dark 12 | 13 | Welcome to Trustee's documentation. Get started with `installation` 14 | and then get an overview with the `quickstart`. The rest of the docs 15 | describe each component of Trustee in detail, with a full reference in 16 | the :doc:`api` section. 17 | 18 | .. raw:: html 19 | 20 | 43 |
44 | 45 | 46 | 47 | 48 | Github Repo 49 | 50 | 51 | 52 | 53 | 54 | Use Cases 55 | 56 | 57 | 58 | 59 | 60 | Tech Report 61 | 62 |
63 | 64 | Overview 65 | ------------- 66 | 67 | Trustee is a framework to extract decision tree explanation from black-box ML models. 68 | 69 | .. figure:: _static/flowchart.png 70 | :align: center 71 | :alt: Trustee Flowchart 72 | 73 | Standard AI/ML development pipeline extended by Trustee. 74 | 75 | 76 | Getting Started 77 | --------------- 78 | This section contains basic information and instructions to get started with Trustee. 79 | 80 | Python Version 81 | *************** 82 | 83 | Trustee supports Python >=3.7. 84 | 85 | Install Trustee 86 | *************** 87 | 88 | Use the following command to install Trustee: 89 | 90 | .. code-block:: sh 91 | 92 | $ pip install trustee 93 | 94 | 95 | Sample Code 96 | ******************* 97 | 98 | .. code:: python 99 | 100 | from sklearn import datasets 101 | from sklearn.ensemble import RandomForestClassifier 102 | from sklearn.model_selection import train_test_split 103 | from sklearn.metrics import classification_report 104 | 105 | from trustee import ClassificationTrustee 106 | 107 | X, y = datasets.load_iris(return_X_y=True) 108 | X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30) 109 | 110 | clf = RandomForestClassifier(n_estimators=100) 111 | clf.fit(X_train, y_train) 112 | y_pred = clf.predict(X_test) 113 | 114 | trustee = ClassificationTrustee(expert=clf) 115 | trustee.fit(X_train, y_train, num_iter=50, num_stability_iter=10, samples_size=0.3, verbose=True) 116 | dt, pruned_dt, agreement, reward = trustee.explain() 117 | dt_y_pred = dt.predict(X_test) 118 | 119 | print("Model explanation global fidelity report:") 120 | print(classification_report(y_pred, dt_y_pred)) 121 | print("Model explanation score report:") 122 | print(classification_report(y_test, dt_y_pred)) 123 | 124 | 125 | Other Use Cases 126 | ******************* 127 | For other examples and use cases of how Trustee can used to scrutinize ML models, listed in the table below, please check our `Use Cases repository `_. 128 | 129 | .. table:: 130 | :class: align-left 131 | 132 | ===================== =========================================================================================================================================================== 133 | Use Case Description 134 | ===================== =========================================================================================================================================================== 135 | `heartbleed_case/` Trustee application to a Random Forest Classifier for an Intrustion Detection System, trained with CIC-IDS-2017 dataset pre-computed features. 136 | `kitsune_case/` Trustee application to Kitsune model for anomaly detection in network traffic, trained with features extracted from Kitsune's Mirai attack trace. 137 | `iot_case/` Trustee application to Random Forest Classifier to distguish IoT devices, trained with features extracted from the pcaps from the UNSW IoT Dataset. 138 | `moon_star_case/` Trustee application to Neural Network Moon and Stars Shortcut learning toy example. 139 | `nprint_ids_case/` Trustee application to the nPrintML AutoGluon Tabular Predictor for an Intrustion Detection System, also trained using pcaps from the CIC-IDS-2017 dataset. 140 | `nprint_os_case/` Trustee application to the nPrintML AutoGluon Tabular Predictor for OS Fingerprinting, also trained using with pcaps from the CIC-IDS-2017 dataset. 141 | `pensieve_case/` Trustee application to the Pensieve RL model for adaptive bit-rate prediction, and comparison to related work Metis. 142 | `vpn_case/` Trustee application the 1D-CNN trained to detect VPN traffic trained with the ISCX VPN-nonVPN dataset. 143 | ===================== =========================================================================================================================================================== 144 | 145 | Supported AI/ML Libraries 146 | ************************* 147 | 148 | .. table:: 149 | :class: align-left 150 | 151 | ============== =================== 152 | Library Supported 153 | ============== =================== 154 | `scikit-learn` |:white_check_mark:| 155 | `Keras` |:white_check_mark:| 156 | `Tensorflow` |:white_check_mark:| 157 | `PyTorch` |:white_check_mark:| 158 | `AutoGluon` |:white_check_mark:| 159 | ============== =================== 160 | 161 | API Reference 162 | ------------- 163 | 164 | If you are looking for information on a specific function, class or 165 | method, this part of the documentation is for you. 166 | 167 | .. toctree:: 168 | :maxdepth: 2 169 | 170 | api 171 | auto_examples/index 172 | 173 | 174 | Citing Us 175 | --------- 176 | 177 | .. code:: 178 | 179 | @inproceedings{Jacobs2022, 180 | title = {AI/ML and Network Security: The Emperor has no Clothes}, 181 | author = {A. S. Jacobs and R. Beltiukov and W. Willinger and R. A. Ferreira and A. Gupta and L. Z. Granville}, 182 | year = 2022, 183 | booktitle = {Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security}, 184 | location = {Los Angeles, CA, USA}, 185 | publisher = {Association for Computing Machinery}, 186 | address = {New York, NY, USA}, 187 | series = {CCS '22} 188 | } 189 | 190 | -------------------------------------------------------------------------------- /docs/make.bat: -------------------------------------------------------------------------------- 1 | @ECHO OFF 2 | 3 | pushd %~dp0 4 | 5 | REM Command file for Sphinx documentation 6 | 7 | if "%SPHINXBUILD%" == "" ( 8 | set SPHINXBUILD=sphinx-build 9 | ) 10 | set SOURCEDIR=source 11 | set BUILDDIR=build 12 | 13 | %SPHINXBUILD% >NUL 2>NUL 14 | if errorlevel 9009 ( 15 | echo. 16 | echo.The 'sphinx-build' command was not found. Make sure you have Sphinx 17 | echo.installed, then set the SPHINXBUILD environment variable to point 18 | echo.to the full path of the 'sphinx-build' executable. Alternatively you 19 | echo.may add the Sphinx directory to PATH. 20 | echo. 21 | echo.If you don't have Sphinx installed, grab it from 22 | echo.https://www.sphinx-doc.org/ 23 | exit /b 1 24 | ) 25 | 26 | if "%1" == "" goto help 27 | 28 | %SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O% 29 | goto end 30 | 31 | :help 32 | %SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O% 33 | 34 | :end 35 | popd 36 | -------------------------------------------------------------------------------- /docs/requirements.txt: -------------------------------------------------------------------------------- 1 | alabaster==0.7.12 ; python_version >= "3.7" 2 | babel==2.10.3 ; python_version >= "3.7" 3 | beautifulsoup4==4.11.1 ; python_version >= "3.7" 4 | certifi==2022.6.15 ; python_version >= "3.7" 5 | charset-normalizer==2.0.12 ; python_version >= "3.7" 6 | colorama==0.4.5 ; sys_platform == "win32" and python_version >= "3.7" 7 | cycler==0.11.0 ; python_version >= "3.7" 8 | docutils==0.19 ; python_version >= "3.7" 9 | fonttools==4.37.1 ; python_version >= "3.7" 10 | furo==2022.6.21 ; python_version >= "3.7" 11 | graphviz==0.20.1 ; python_version >= "3.7" 12 | idna==3.3 ; python_version >= "3.7" 13 | imagesize==1.4.1 ; python_version >= "3.7" 14 | importlib-metadata==4.12.0 ; python_version >= "3.7" and python_version < "3.10" 15 | jinja2==3.1.2 ; python_version >= "3.7" 16 | joblib==1.1.0 ; python_version >= "3.7" 17 | kiwisolver==1.4.4 ; python_version >= "3.7" 18 | markupsafe==2.1.1 ; python_version >= "3.7" 19 | matplotlib==3.5.3 ; python_version >= "3.7" 20 | numpy==1.21.1 ; python_version >= "3.7" 21 | packaging==21.3 ; python_version >= "3.7" 22 | pandas==1.1.5 ; python_version >= "3.7" 23 | pillow==9.2.0 ; python_version >= "3.7" 24 | prettytable==3.0.0 ; python_version >= "3.7" 25 | pygments==2.13.0 ; python_version >= "3.7" 26 | pyparsing==3.0.9 ; python_version >= "3.7" 27 | python-dateutil==2.8.2 ; python_version >= "3.7" 28 | pytz==2022.2.1 ; python_version >= "3.7" 29 | requests==2.27.1 ; python_version >= "3.7" 30 | scikit-learn==1.0.2 ; python_version >= "3.7" 31 | scipy==1.6.1 ; python_version >= "3.7" 32 | setuptools-scm==6.4.2 ; python_version >= "3.7" 33 | setuptools==57.5.0 ; python_version >= "3.7" 34 | six==1.16.0 ; python_version >= "3.7" 35 | snowballstemmer==2.2.0 ; python_version >= "3.7" 36 | soupsieve==2.3.2.post1 ; python_version >= "3.7" 37 | sphinx-basic-ng==0.0.1a12 ; python_version >= "3.7" 38 | sphinx-gallery==0.11.1 ; python_version >= "3.7" 39 | sphinx==5.1.1 ; python_version >= "3.7" 40 | sphinxcontrib-applehelp==1.0.2 ; python_version >= "3.7" 41 | sphinxcontrib-devhelp==1.0.2 ; python_version >= "3.7" 42 | sphinxcontrib-htmlhelp==2.0.0 ; python_version >= "3.7" 43 | sphinxcontrib-jsmath==1.0.1 ; python_version >= "3.7" 44 | sphinxcontrib-qthelp==1.0.3 ; python_version >= "3.7" 45 | sphinxcontrib-serializinghtml==1.1.5 ; python_version >= "3.7" 46 | sphinxemoji==0.2.0 ; python_version >= "3.7" 47 | termcolor==1.1.0 ; python_version >= "3.7" 48 | threadpoolctl==3.1.0 ; python_version >= "3.7" 49 | tomli==2.0.1 ; python_version >= "3.7" 50 | typing-extensions==4.3.0 ; python_version < "3.8" and python_version >= "3.7" 51 | urllib3==1.22 ; python_version >= "3.7" 52 | wcwidth==0.2.5 ; python_version >= "3.7" 53 | zipp==3.8.1 ; python_version >= "3.7" and python_version < "3.10" 54 | -------------------------------------------------------------------------------- /examples/README.rst: -------------------------------------------------------------------------------- 1 | .. _examples-index: 2 | 3 | Examples 4 | ================== 5 | 6 | Below is a gallery of examples of how to use the Trustee module. -------------------------------------------------------------------------------- /examples/classification.py: -------------------------------------------------------------------------------- 1 | """ 2 | ClassificationTrustee 3 | ===================== 4 | 5 | Simple example on how to use the ClassificationTrustee class to extract 6 | a decision tree from a RandomForestClassifier from scikit-learn. 7 | """ 8 | # importing required libraries 9 | # importing Scikit-learn library and datasets package 10 | import graphviz 11 | 12 | from sklearn import tree 13 | from sklearn import datasets 14 | from sklearn.ensemble import RandomForestClassifier 15 | from sklearn.model_selection import train_test_split 16 | from sklearn.metrics import classification_report 17 | 18 | 19 | from trustee import ClassificationTrustee 20 | 21 | # Loading the iris plants dataset (classification) 22 | iris = datasets.load_iris() 23 | X, y = datasets.load_iris(return_X_y=True) 24 | 25 | # Spliting arrays or matrices into random train and test subsets 26 | # i.e. 70 % training dataset and 30 % test datasets 27 | X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30) 28 | 29 | # creating a RF classifier 30 | clf = RandomForestClassifier(n_estimators=100) 31 | # Training the model on the training dataset 32 | # fit function is used to train the model using the training sets as parameters 33 | clf.fit(X_train, y_train) 34 | # performing predictions on the test dataset 35 | y_pred = clf.predict(X_test) 36 | 37 | # Evaluate model accuracy 38 | print("Model classification report:") 39 | print(classification_report(y_test, y_pred)) 40 | 41 | # Initialize Trustee and fit for classification models 42 | trustee = ClassificationTrustee(expert=clf) 43 | trustee.fit(X_train, y_train, num_iter=50, num_stability_iter=10, samples_size=0.3, verbose=True) 44 | 45 | # Get the best explanation from Trustee 46 | dt, pruned_dt, agreement, reward = trustee.explain() 47 | print(f"Model explanation training (agreement, fidelity): ({agreement}, {reward})") 48 | print(f"Model Explanation size: {dt.tree_.node_count}") 49 | print(f"Top-k Prunned Model explanation size: {pruned_dt.tree_.node_count}") 50 | 51 | # Use explanations to make predictions 52 | dt_y_pred = dt.predict(X_test) 53 | pruned_dt_y_pred = pruned_dt.predict(X_test) 54 | 55 | # Evaluate accuracy and fidelity of explanations 56 | print("Model explanation global fidelity report:") 57 | print(classification_report(y_pred, dt_y_pred)) 58 | print("Top-k Model explanation global fidelity report:") 59 | print(classification_report(y_pred, pruned_dt_y_pred)) 60 | 61 | print("Model explanation score report:") 62 | print(classification_report(y_test, dt_y_pred)) 63 | print("Top-k Model explanation score report:") 64 | print(classification_report(y_test, pruned_dt_y_pred)) 65 | 66 | 67 | # Output decision tree to pdf 68 | dot_data = tree.export_graphviz( 69 | dt, 70 | class_names=iris.target_names, 71 | feature_names=iris.feature_names, 72 | filled=True, 73 | rounded=True, 74 | special_characters=True, 75 | ) 76 | graph = graphviz.Source(dot_data) 77 | graph.render("dt_explanation") 78 | 79 | # Output pruned decision tree to pdf 80 | dot_data = tree.export_graphviz( 81 | pruned_dt, 82 | class_names=iris.target_names, 83 | feature_names=iris.feature_names, 84 | filled=True, 85 | rounded=True, 86 | special_characters=True, 87 | ) 88 | graph = graphviz.Source(dot_data) 89 | graph.render("pruned_dt_explation") 90 | -------------------------------------------------------------------------------- /examples/regression.py: -------------------------------------------------------------------------------- 1 | """ 2 | RegressionTrustee 3 | ================= 4 | 5 | Simple example on how to use the ClassificationTrustee class to extract 6 | a decision tree from a MLPRegressor (neural network) from scikit-learn. 7 | """ 8 | # importing required libraries 9 | # importing Scikit-learn library and datasets package 10 | import graphviz 11 | 12 | from sklearn import tree 13 | from sklearn import datasets 14 | from sklearn.neural_network import MLPRegressor 15 | from sklearn.model_selection import train_test_split 16 | from sklearn.metrics import r2_score 17 | 18 | from trustee import RegressionTrustee 19 | 20 | # Loading the diabetes dataset (regression) 21 | diabetes = datasets.load_diabetes() 22 | X, y = datasets.load_diabetes(return_X_y=True) 23 | # Spliting arrays or matrices into random train and test subsets 24 | # i.e. 70 % training dataset and 30 % test datasets 25 | X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30) 26 | 27 | # creating a MLP regressor 28 | clf = MLPRegressor(solver="lbfgs", alpha=1e-5, hidden_layer_sizes=(100, 50), max_iter=500, random_state=1) 29 | # Training the model on the training dataset 30 | # fit function is used to train the model using the training sets as parameters 31 | clf.fit(X_train, y_train) 32 | # performing predictions on the test dataset 33 | y_pred = clf.predict(X_test) 34 | 35 | # Evaluate model accuracy 36 | print("Model R2-score:") 37 | print(r2_score(y_test, y_pred)) 38 | 39 | # Initialize Trustee and fit for classification models 40 | trustee = RegressionTrustee(expert=clf) 41 | trustee.fit(X_train, y_train, num_iter=50, num_stability_iter=10, samples_size=0.3, verbose=True) 42 | 43 | # Get the best explanation from Trustee 44 | dt, pruned_dt, agreement, reward = trustee.explain() 45 | print(f"Model explanation training (agreement, fidelity): ({agreement}, {reward})") 46 | print(f"Model Explanation size: {dt.tree_.node_count}") 47 | print(f"Top-k Prunned Model explanation size: {pruned_dt.tree_.node_count}") 48 | 49 | # Use explanations to make predictions 50 | dt_y_pred = dt.predict(X_test) 51 | pruned_dt_y_pred = pruned_dt.predict(X_test) 52 | 53 | # Evaluate accuracy and fidelity of explanations 54 | print("Model explanation global fidelity:") 55 | print(r2_score(y_pred, dt_y_pred)) 56 | print("Top-k Model explanation global fidelity:") 57 | print(r2_score(y_pred, pruned_dt_y_pred)) 58 | 59 | print("Model explanation R2-score:") 60 | print(r2_score(y_test, dt_y_pred)) 61 | print("Top-k Model explanation R2-score:") 62 | print(r2_score(y_test, pruned_dt_y_pred)) 63 | 64 | # Output decision tree to pdf 65 | dot_data = tree.export_graphviz( 66 | dt, 67 | feature_names=diabetes.feature_names, 68 | filled=True, 69 | rounded=True, 70 | special_characters=True, 71 | ) 72 | graph = graphviz.Source(dot_data) 73 | graph.render("dt_explanation") 74 | 75 | # Output pruned decision tree to pdf 76 | dot_data = tree.export_graphviz( 77 | pruned_dt, 78 | feature_names=diabetes.feature_names, 79 | filled=True, 80 | rounded=True, 81 | special_characters=True, 82 | ) 83 | graph = graphviz.Source(dot_data) 84 | graph.render("pruned_dt_explation") 85 | -------------------------------------------------------------------------------- /examples/trust_classification.py: -------------------------------------------------------------------------------- 1 | """ 2 | TrustReport for Classification 3 | ============================== 4 | 5 | Simple example on how to use the TrustReport class to analyze the explanations 6 | produced by ClassificationTrustee from a RandomForestClassifier from scikit-learn. 7 | Notice that using the method `TrustReport.load()`, one can load a previously 8 | generated report saved using `trust_report.save()`. 9 | """ 10 | import os 11 | 12 | # importing required libraries 13 | # importing Scikit-learn library and datasets package 14 | from sklearn import datasets 15 | from sklearn.ensemble import RandomForestClassifier 16 | 17 | from trustee.report.trust import TrustReport 18 | 19 | OUTPUT_PATH = "out/" 20 | REPORT_PATH = f"{OUTPUT_PATH}/report/trust_report.obj" 21 | 22 | if os.path.exists(REPORT_PATH): 23 | print(f"Loading trust report from {REPORT_PATH}...") 24 | trust_report = TrustReport.load(REPORT_PATH) 25 | print("Done!") 26 | else: 27 | # Loading the iris plants dataset (classification) 28 | iris = datasets.load_iris() 29 | # dividing the datasets into two parts i.e. training datasets and test datasets 30 | X, y = datasets.load_iris(return_X_y=True, as_frame=True) 31 | 32 | # creating a RF classifier 33 | clf = RandomForestClassifier(n_estimators=100) 34 | 35 | # The trust report (can) fit and explain the classifier 36 | trust_report = TrustReport( 37 | clf, 38 | X=X, 39 | y=y, 40 | max_iter=5, 41 | num_pruning_iter=5, 42 | train_size=0.7, 43 | trustee_num_iter=10, 44 | trustee_num_stability_iter=5, 45 | trustee_sample_size=0.3, 46 | analyze_branches=True, 47 | analyze_stability=True, 48 | top_k=10, 49 | verbose=True, 50 | class_names=iris.target_names, 51 | feature_names=iris.feature_names, 52 | is_classify=True, 53 | ) 54 | 55 | print(trust_report) 56 | trust_report.save(OUTPUT_PATH) 57 | -------------------------------------------------------------------------------- /examples/trust_regression.py: -------------------------------------------------------------------------------- 1 | """ 2 | TrustReport for Regression 3 | ========================== 4 | 5 | Simple example on how to use the TrustReport class to analyze the explanations 6 | produced by RegressionTrustee from a MLPRegressor (neural network) from scikit-learn. 7 | Notice that using the method `TrustReport.load()`, one can load a previously 8 | generated report saved using `trust_report.save()`. 9 | """ 10 | import os 11 | 12 | # importing required libraries 13 | # importing Scikit-learn library and datasets package 14 | from sklearn import datasets 15 | from sklearn.neural_network import MLPRegressor 16 | 17 | from trustee.report.trust import TrustReport 18 | 19 | OUTPUT_PATH = "out/" 20 | REPORT_PATH = f"{OUTPUT_PATH}/report/trust_report.obj" 21 | 22 | if os.path.exists(REPORT_PATH): 23 | print(f"Loading trust report from {REPORT_PATH}...") 24 | trust_report = TrustReport.load(REPORT_PATH) 25 | print("Done!") 26 | else: 27 | # Loading the diabetes dataset (regression) 28 | diabetes = datasets.load_diabetes() 29 | X, y = datasets.load_diabetes(return_X_y=True, as_frame=True) 30 | 31 | # creating a RF classifier 32 | clf = MLPRegressor(solver="lbfgs", alpha=1e-5, hidden_layer_sizes=(100, 50), max_iter=500, random_state=1) 33 | 34 | # The trust report (can) fit and explain the classifier 35 | trust_report = TrustReport( 36 | clf, 37 | X=X, 38 | y=y, 39 | max_iter=5, 40 | num_pruning_iter=5, 41 | train_size=0.7, 42 | trustee_num_iter=10, 43 | trustee_num_stability_iter=5, 44 | trustee_sample_size=0.3, 45 | analyze_branches=True, 46 | analyze_stability=True, 47 | top_k=10, 48 | verbose=True, 49 | feature_names=diabetes.feature_names, 50 | is_classify=False, # <----- to run the trust report for a regression model 51 | ) 52 | 53 | print(trust_report) 54 | trust_report.save(OUTPUT_PATH) 55 | -------------------------------------------------------------------------------- /pyproject.toml: -------------------------------------------------------------------------------- 1 | [tool.poetry] 2 | name = "trustee" 3 | version = "1.1.6" 4 | readme = "README.md" 5 | description = "This package implements the Trustee framework to extract decision tree explanation from black-box ML models." 6 | homepage = "https://trusteeml.github.io" 7 | repository = "https://github.com/TrusteeML/trustee" 8 | authors = ["Arthur Selle Jacobs "] 9 | packages = [ { include = "trustee"} ] 10 | classifiers = [ 11 | "Intended Audience :: Science/Research", 12 | "Intended Audience :: Developers", 13 | "Topic :: Software Development", 14 | "Topic :: Scientific/Engineering", 15 | "Topic :: Scientific/Engineering :: Artificial Intelligence", 16 | "Topic :: Software Development :: Libraries :: Python Modules", 17 | ] 18 | 19 | [tool.poetry.dependencies] 20 | python = ">=3.7" 21 | numpy = ">=1.19.0" 22 | scipy = "^1.4.1" 23 | pandas = "^1.1.0" 24 | scikit-learn = ">=0.23.2" 25 | matplotlib = "^3.3.1" 26 | setuptools = "^57.0.0" 27 | prettytable = "3.0.0" 28 | termcolor = "^1.1.0" 29 | graphviz = ">=0.8.1" 30 | furo = "^2022.6.21" 31 | sphinxemoji = "^0.2.0" 32 | sphinx-gallery = "^0.11.1" 33 | 34 | [tool.poetry.dev-dependencies] 35 | Sphinx = "^5.1.1" 36 | 37 | [tool.black] 38 | line-length = 120 39 | 40 | [build-system] 41 | requires = ["poetry>=0.12", "setuptools>=^57.0.0"] 42 | build-backend = "poetry.masonry.api" 43 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | alabaster==0.7.12 ; python_version >= "3.7" 2 | babel==2.10.3 ; python_version >= "3.7" 3 | beautifulsoup4==4.11.1 ; python_version >= "3.7" 4 | certifi==2022.6.15 ; python_version >= "3.7" 5 | charset-normalizer==2.0.12 ; python_version >= "3.7" 6 | colorama==0.4.5 ; sys_platform == "win32" and python_version >= "3.7" 7 | cycler==0.11.0 ; python_version >= "3.7" 8 | docutils==0.19 ; python_version >= "3.7" 9 | fonttools==4.37.1 ; python_version >= "3.7" 10 | furo==2022.6.21 ; python_version >= "3.7" 11 | graphviz==0.20.1 ; python_version >= "3.7" 12 | idna==3.3 ; python_version >= "3.7" 13 | imagesize==1.4.1 ; python_version >= "3.7" 14 | importlib-metadata==4.12.0 ; python_version >= "3.7" and python_version < "3.10" 15 | jinja2==3.1.2 ; python_version >= "3.7" 16 | joblib==1.1.0 ; python_version >= "3.7" 17 | kiwisolver==1.4.4 ; python_version >= "3.7" 18 | markupsafe==2.1.1 ; python_version >= "3.7" 19 | matplotlib==3.5.3 ; python_version >= "3.7" 20 | numpy==1.21.1 ; python_version >= "3.7" 21 | packaging==21.3 ; python_version >= "3.7" 22 | pandas==1.1.5 ; python_version >= "3.7" 23 | pillow==9.2.0 ; python_version >= "3.7" 24 | prettytable==3.0.0 ; python_version >= "3.7" 25 | pygments==2.13.0 ; python_version >= "3.7" 26 | pyparsing==3.0.9 ; python_version >= "3.7" 27 | python-dateutil==2.8.2 ; python_version >= "3.7" 28 | pytz==2022.2.1 ; python_version >= "3.7" 29 | requests==2.27.1 ; python_version >= "3.7" 30 | scikit-learn==1.0.2 ; python_version >= "3.7" 31 | scipy==1.6.1 ; python_version >= "3.7" 32 | setuptools-scm==6.4.2 ; python_version >= "3.7" 33 | setuptools==57.5.0 ; python_version >= "3.7" 34 | six==1.16.0 ; python_version >= "3.7" 35 | snowballstemmer==2.2.0 ; python_version >= "3.7" 36 | soupsieve==2.3.2.post1 ; python_version >= "3.7" 37 | sphinx-basic-ng==0.0.1a12 ; python_version >= "3.7" 38 | sphinx-gallery==0.11.1 ; python_version >= "3.7" 39 | sphinx==5.1.1 ; python_version >= "3.7" 40 | sphinxcontrib-applehelp==1.0.2 ; python_version >= "3.7" 41 | sphinxcontrib-devhelp==1.0.2 ; python_version >= "3.7" 42 | sphinxcontrib-htmlhelp==2.0.0 ; python_version >= "3.7" 43 | sphinxcontrib-jsmath==1.0.1 ; python_version >= "3.7" 44 | sphinxcontrib-qthelp==1.0.3 ; python_version >= "3.7" 45 | sphinxcontrib-serializinghtml==1.1.5 ; python_version >= "3.7" 46 | sphinxemoji==0.2.0 ; python_version >= "3.7" 47 | termcolor==1.1.0 ; python_version >= "3.7" 48 | threadpoolctl==3.1.0 ; python_version >= "3.7" 49 | tomli==2.0.1 ; python_version >= "3.7" 50 | typing-extensions==4.3.0 ; python_version < "3.8" and python_version >= "3.7" 51 | urllib3==1.22 ; python_version >= "3.7" 52 | wcwidth==0.2.5 ; python_version >= "3.7" 53 | zipp==3.8.1 ; python_version >= "3.7" and python_version < "3.10" 54 | -------------------------------------------------------------------------------- /trustee/__init__.py: -------------------------------------------------------------------------------- 1 | from .main import * 2 | from ._version import __version__ 3 | -------------------------------------------------------------------------------- /trustee/_version.py: -------------------------------------------------------------------------------- 1 | __version__ = "1.1.6" 2 | -------------------------------------------------------------------------------- /trustee/enums/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TrusteeML/trustee/71561fbcc96af8596fbd23879c3edcf0639e985b/trustee/enums/__init__.py -------------------------------------------------------------------------------- /trustee/enums/feature_type.py: -------------------------------------------------------------------------------- 1 | """ Feature Type Enum """ 2 | from enum import Enum 3 | 4 | 5 | class FeatureType(Enum): 6 | """ Types of features in training/testing datasets """ 7 | CATEGORICAL = 0 8 | NUMERICAL = 1 9 | IDENTIFIER = 2 10 | -------------------------------------------------------------------------------- /trustee/main.py: -------------------------------------------------------------------------------- 1 | """ 2 | Trustee 3 | ==================================== 4 | The core module of the Trustee project 5 | """ 6 | import abc 7 | import functools 8 | import numpy as np 9 | import pandas as pd 10 | 11 | from copy import deepcopy 12 | 13 | 14 | from sklearn.metrics import f1_score, r2_score 15 | from sklearn.model_selection import train_test_split 16 | from sklearn.tree import DecisionTreeClassifier, DecisionTreeRegressor 17 | 18 | from trustee.utils.tree import get_dt_info, top_k_prune 19 | from trustee.utils.dataset import convert_to_df, convert_to_series 20 | 21 | 22 | def _check_if_trained(func): 23 | """ 24 | Checks whether the Trustee is already fitted and self._best_student exists 25 | 26 | Parameters 27 | ---------- 28 | func: callable 29 | Function to apply decorator to. 30 | *args: tuple 31 | Additional arguments should be passed as keyword arguments to `func`. 32 | **kwargs: dict, optional 33 | Extra arguments to `func`: refer to each func documentation for a list of all possible arguments. 34 | """ 35 | 36 | @functools.wraps(func) 37 | def wrapper(self, *args, **kwargs): 38 | if len(self._top_students) == 0: 39 | raise ValueError("No student models have been trained yet. Please fit() Trustee explainer first.") 40 | return func(self, *args, **kwargs) 41 | 42 | return wrapper 43 | 44 | 45 | class Trustee(abc.ABC): 46 | """ 47 | Base implementation the Trust-oriented Decision Tree Extraction (Trustee) 48 | algorithm to train student model based on observations from an Expert model. 49 | """ 50 | 51 | def __init__(self, expert, student_class, logger=None): 52 | """ 53 | Trustee constructor. 54 | 55 | Parameters 56 | ---------- 57 | expert: object 58 | The ML blackbox model to analyze. The expert model must have a `predict` method call implemented for 59 | Trustee to work properly, unless explicitly stated otherwise using the `predict_method_name` argument 60 | in the fit() method. 61 | 62 | student_class: Class 63 | Class of student to train based on blackbox model predictions. The given Class must implement a `fit() 64 | and a `predict()` method interface for Trustee to work properly. The current implementation has been 65 | tested using the DecisionTreeClassifier and DecisionTreeRegressor from scikit-learn. 66 | 67 | logger: Logger object , default=None 68 | A logger object to log messages to. If none is given, the print() method will be used to log messages. 69 | """ 70 | self.log = logger.log if logger else print 71 | self.expert = expert 72 | self.student_class = student_class 73 | 74 | self._students_by_iter = [] 75 | self._top_students = [] 76 | self._stable_students = [] 77 | 78 | self._X_train = [] 79 | self._X_test = [] 80 | self._y_train = [] 81 | self._y_test = [] 82 | 83 | self._best_student = None 84 | self._features = None 85 | self._nodes = None 86 | self._branches = None 87 | 88 | self._student_use_features: np.array = [] 89 | 90 | @abc.abstractmethod 91 | def _score(self, y_true, y_pred): 92 | """ 93 | Score function for student models. Compares the ground-truth predictions 94 | of a blackbox model with the predictions of a student model. 95 | 96 | Parameters 97 | ---------- 98 | y_true: array-like of shape (n_samples,) or (n_samples, n_outputs) 99 | The ground-truth target values (class labels in classification, real numbers in regression). 100 | 101 | y_pred: array-like of shape (n_samples,) or (n_samples, n_outputs) 102 | The predicted target values (class labels in classification, real numbers in regression). 103 | 104 | Returns 105 | ------- 106 | score: float 107 | Calculated student model score. 108 | """ 109 | 110 | def fit( 111 | self, 112 | X, 113 | y, 114 | top_k=10, 115 | max_leaf_nodes=None, 116 | max_depth=None, 117 | ccp_alpha=0.0, 118 | train_size=0.7, 119 | num_iter=50, 120 | num_stability_iter=5, 121 | num_samples=2000, 122 | samples_size=None, 123 | use_features=None, 124 | predict_method_name="predict", 125 | optimization="fidelity", # for comparative purposes only 126 | aggregate=True, # for comparative purposes only 127 | verbose=False, 128 | ): 129 | """ 130 | Trains Decision Tree Regressor to imitate Expert model. 131 | 132 | Parameters 133 | ---------- 134 | X: {array-like, sparse matrix} of shape (n_samples, n_features) 135 | The training input samples. Internally, it will be converted to a pandas DataFrame. 136 | 137 | y: array-like of shape (n_samples,) or (n_samples, n_outputs) 138 | The target values for X (class labels in classification, real numbers in regression). 139 | Internally, it will be converted to a pandas Series. 140 | 141 | top_k: int, default=10 142 | Number of top-k branches, sorted by number of samples per branch, to keep after finding 143 | decision tree with highest fidelity. 144 | 145 | max_leaf_nodes: int, default=None 146 | Grow a tree with max_leaf_nodes in best-first fashion. Best nodes are defined as 147 | relative reduction in impurity. If None then unlimited number of leaf nodes. 148 | 149 | max_depth: int, default=None 150 | The maximum depth of the tree. If None, then nodes are expanded until all leaves are pure. 151 | 152 | ccp_alpha: float, default=0.0 153 | Complexity parameter used for Minimal Cost-Complexity Pruning. The subtree with the 154 | largest cost complexity that is smaller than ccp_alpha will be chosen. By default, 155 | no pruning is performed. See Minimal Cost-Complexity Pruning here for details: 156 | https://scikit-learn.org/stable/modules/tree.html#minimal-cost-complexity-pruning 157 | 158 | train_size: float or int, default=0.7 159 | If float, should be between 0.0 and 1.0 and represent the proportion of the dataset 160 | to include in the train split. If int, represents the absolute number of train samples. 161 | 162 | num_iter: int, default=50 163 | Number of iterations to repeat Trustee inner-loop for. 164 | 165 | num_stability_iter: int, default=5 166 | Number of stability to repeat Trustee stabilization outer-loop for. 167 | 168 | num_samples: int, default=2000 169 | The absolute number of samples to fetch from the training dataset split to train the 170 | student decision tree model. If the `samples_size` argument is provided, this arg is 171 | ignored. 172 | 173 | samples_size: float, default=None 174 | The fraction of the training dataset to use to train the student decision tree model. 175 | If None, the value is automatically set to the `num_samples` provided value. 176 | 177 | use_features: array-like, default=None 178 | Array-like of integers representing the indexes of features from the `X` training samples. 179 | If not None, only the features indicated by the provided indexes will be used to train the 180 | student decision tree model. 181 | 182 | predict_method_name: str, default="predict" 183 | The method interface to use to get predictions from the expert model. 184 | If no value is passed, the default `predict` interface is used. 185 | 186 | optimization: {"fidelity", "accuracy"}, default="fidelity" 187 | The comparison criteria to optimize the decision tree students in Trustee inner-loop. 188 | Used for ablation study only. 189 | 190 | aggregate: bool, default=True 191 | Boolean indicating whether dataset aggregation should be used in Trustee inner-loop. 192 | Used for ablation study only. 193 | 194 | verbose: bool, default=False 195 | Boolean indicating whether to log messages. 196 | """ 197 | if verbose: 198 | self.log(f"Initializing training dataset using {self.expert} as expert model") 199 | 200 | if len(X) != len(y): 201 | raise ValueError("Features (X) and target (y) values should have the same length.") 202 | 203 | # convert data to np array to facilitate processing 204 | X = convert_to_df(X) 205 | y = convert_to_series(y) 206 | 207 | self._student_use_features = use_features if use_features is not None else np.arange(0, len(X.columns)) 208 | 209 | # split input array to train DTs and evaluate agreement 210 | self._X_train, self._X_test, self._y_train, self._y_test = train_test_split(X, y, train_size=train_size) 211 | 212 | features = self._X_train 213 | targets = convert_to_series(getattr(self.expert, predict_method_name)(self._X_train)) 214 | 215 | if hasattr(targets, "shape") and len(targets.shape) >= 2: 216 | targets = targets.ravel() 217 | 218 | student = self.student_class( 219 | random_state=0, max_leaf_nodes=max_leaf_nodes, max_depth=max_depth, ccp_alpha=ccp_alpha 220 | ) 221 | 222 | if verbose: 223 | self.log(f"Expert model score: {self._score(self._y_train, targets)}") 224 | self.log(f"Initializing Trustee outer-loop with {num_stability_iter} iterations") 225 | 226 | # Trustee outer-loop 227 | for i in range(num_stability_iter): 228 | self._students_by_iter.append([]) 229 | if verbose: 230 | self.log("#" * 10, f"Outer-loop Iteration {i}/{num_stability_iter}", "#" * 10) 231 | self.log(f"Initializing Trustee inner-loop with {num_stability_iter} iterations") 232 | 233 | # Trustee inner-loop 234 | for j in range(num_iter): 235 | if verbose: 236 | self.log("#" * 10, f"Inner-loop Iteration {j}/{num_iter}", "#" * 10) 237 | 238 | dataset_size = len(features) 239 | size = int(int(len(self._X_train)) * samples_size) if samples_size else num_samples 240 | # Step 1: Sample predictions from training dataset 241 | if verbose: 242 | self.log( 243 | f"Sampling {size} points from training dataset with ({len(features)}, {len(targets)}) entries" 244 | ) 245 | 246 | samples_idxs = np.random.choice(dataset_size, size=size, replace=False) 247 | X_iter, y_iter = features.iloc[samples_idxs], targets.iloc[samples_idxs] 248 | X_iter_train, X_iter_test, y_iter_train, y_iter_test = train_test_split( 249 | X_iter, y_iter, train_size=train_size 250 | ) 251 | X_train_student = X_iter_train.iloc[:, self._student_use_features] 252 | X_test_student = X_iter_test.iloc[:, self._student_use_features] 253 | 254 | # Step 2: Training DecisionTreeRegressor with sampled data 255 | student.fit(X_train_student.values, y_iter_train.values) 256 | student_pred = student.predict(X_test_student.values) 257 | 258 | if verbose: 259 | self.log( 260 | f"Student model {i}-{j} trained with depth {student.get_depth()} " 261 | f"and {student.get_n_leaves()} leaves:" 262 | ) 263 | self.log(f"Student model score: {self._score(y_iter_test, student_pred)}") 264 | 265 | # Step 3: Use expert model predictions to aggregate original dataset 266 | expert_pred = pd.Series(getattr(self.expert, predict_method_name)(X_iter_test)) 267 | if hasattr(expert_pred, "shape") and len(expert_pred.shape) >= 2: 268 | expert_pred = expert_pred.ravel() 269 | 270 | if aggregate: 271 | features = pd.concat([features, X_iter_test]) 272 | targets = pd.concat([targets, expert_pred]) 273 | 274 | if optimization == "accuracy": 275 | # Step 4: Calculate reward based on Decision Tree Classifier accuracy 276 | reward = self._score(y_iter_test, student_pred) 277 | else: 278 | # Step 4: Calculate reward based on Decision Tree Classifier fidelity to the Expert model 279 | reward = self._score(expert_pred, student_pred) 280 | 281 | if verbose: 282 | self.log(f"Student model {i}-{j} fidelity: {reward}") 283 | 284 | # Save student to list of iterations dt 285 | self._students_by_iter[i].append((deepcopy(student), reward)) 286 | 287 | # Save student with highest fidelity to list of top students by iteration 288 | self._top_students.append(max(self._students_by_iter[i], key=lambda item: item[1])) 289 | 290 | # Get best overall student based on mean agreement 291 | self._best_student = self.explain(top_k=top_k)[0] 292 | 293 | @_check_if_trained 294 | def explain(self, top_k=10): 295 | """ 296 | Returns explainable model that best imitates Expert model, based on highest mean agreement and highest fidelity. 297 | 298 | Returns 299 | ------- 300 | top_student: tuple 301 | (dt, pruned_dt, agreement, reward) 302 | 303 | - dt: {DecisionTreeClassifier, DecisionTreeRegressor} 304 | Unconstrained fitted student model. 305 | 306 | - pruned_dt: {DecisionTreeClassifier, DecisionTreeRegressor} 307 | Top-k pruned fitted student model. 308 | 309 | - agreement: float 310 | Mean agreement of pruned student model with respect to others. 311 | 312 | - reward: float 313 | Fidelity of student model to the expert model. 314 | """ 315 | stable = self.get_stable(top_k=top_k, threshold=0, sort=False) 316 | return max(stable, key=lambda item: item[2]) 317 | 318 | @_check_if_trained 319 | def get_stable(self, top_k=10, threshold=0.9, sort=True): 320 | """ 321 | Filters out explanations from Trustee stability analysis with less than threshold agreement. 322 | 323 | Parameters 324 | ---------- 325 | top_k: int, default=10 326 | Number of top-k branches, sorted by number of samples per branch, to keep after finding 327 | decision tree with highest fidelity. 328 | 329 | threshold: float, default=0.9 330 | Remove any student decision tree explanation if their mean agreement goes below given threshold. 331 | To keep all students regardless of mean agreement, pass 0. 332 | 333 | sort: bool, default=True 334 | Boolean indicating whether to sort returned stable student explanation based on mean agreement. 335 | 336 | Returns 337 | ------- 338 | stable_explanations: array-like of tuple 339 | [(dt, pruned_dt, agreement, reward), ...] 340 | 341 | - dt: {DecisionTreeClassifier, DecisionTreeRegressor} 342 | Unconstrained fitted student model. 343 | 344 | - pruned_dt: {DecisionTreeClassifier, DecisionTreeRegressor} 345 | Top-k pruned fitted student model. 346 | 347 | - agreement: float 348 | Mean agreement of pruned student model with respect to others. 349 | 350 | - reward: float 351 | Fidelity of student model to the expert model. 352 | """ 353 | if len(self._stable_students) == 0: 354 | agreement = [] 355 | # Calculate pair-wise agreement of all top students generated during inner loop 356 | for i, _ in enumerate(self._top_students): 357 | agreement.append([]) 358 | # Apply top-k pruning before calculating agreement 359 | base_tree = top_k_prune(self._top_students[i][0], top_k=top_k) 360 | for j, _ in enumerate(self._top_students): 361 | # Apply top-k pruning before calculating agreement 362 | iter_tree = top_k_prune(self._top_students[j][0], top_k=top_k) 363 | 364 | iter_y_pred = iter_tree.predict(self._X_test.iloc[:, self._student_use_features].values) 365 | base_y_pred = base_tree.predict(self._X_test.iloc[:, self._student_use_features].values) 366 | 367 | agreement[i].append(self._score(iter_y_pred, base_y_pred)) 368 | 369 | # Save complete dt, top-k prune dt, mean agreement and fidelity 370 | self._stable_students.append( 371 | ( 372 | self._top_students[i][0], 373 | base_tree, 374 | np.mean(agreement[i]), 375 | self._top_students[i][1], 376 | ) 377 | ) 378 | 379 | stable = self._stable_students 380 | if threshold > 0: 381 | stable = filter(lambda item: item[2] >= threshold, stable) 382 | 383 | if sort: 384 | return sorted(stable, key=lambda item: item[2], reverse=True) 385 | 386 | return stable 387 | 388 | @_check_if_trained 389 | def get_all_students(self): 390 | """ 391 | Get list of all (student, reward) obtained during the inner-loop process. 392 | 393 | Returns 394 | ------- 395 | students_by_iter: array-like of shape (num_stability_iter, num_iter) of tuple (dt, reward) 396 | Matrix with all student models trained during `fit()`. 397 | 398 | - dt: {DecisionTreeClassifier, DecisionTreeRegressor} 399 | Unconstrained fitted student model. 400 | 401 | - reward: float 402 | Fidelity of student model to the expert model. 403 | """ 404 | return self._students_by_iter 405 | 406 | @_check_if_trained 407 | def get_top_students(self): 408 | """ 409 | Get list of top (students, reward) obtained during the outer-loop process. 410 | 411 | Returns 412 | ------- 413 | top_students: array-like of shape (num_stability_iter,) of tuple (dt, reward) 414 | List with top student models trained during `fit()`. 415 | 416 | - dt: {DecisionTreeClassifier, DecisionTreeRegressor} 417 | Unconstrained fitted student model. 418 | 419 | - reward: float 420 | Fidelity of student model to the expert model. 421 | """ 422 | return self._top_students 423 | 424 | @_check_if_trained 425 | def get_n_features(self): 426 | """ 427 | Returns number of features used in the top student model. 428 | 429 | Returns 430 | ------- 431 | n_features: int 432 | Number of features used in top student model. 433 | """ 434 | if not self._features: 435 | self._features, self._nodes, self._branches = get_dt_info(self._best_student) 436 | 437 | return len(self._features.keys()) 438 | 439 | @_check_if_trained 440 | def get_n_classes(self): 441 | """ 442 | Returns number of classes used in the top student model. 443 | 444 | Returns 445 | ------- 446 | n_classes: int 447 | Number of classes outputted in top student model. 448 | """ 449 | return self._best_student.tree_.n_classes[0] 450 | 451 | @_check_if_trained 452 | def get_samples_sum(self): 453 | """ 454 | Get the sum of all samples in all non-leaf _nodes in best student model. 455 | 456 | Returns 457 | ------- 458 | samples_sum: int 459 | Sum of all samples covered by non-leaf nodes in top student model. 460 | """ 461 | left = self._best_student.tree_.children_left 462 | right = self._best_student.tree_.children_right 463 | samples = self._best_student.tree_.n_node_samples 464 | 465 | return np.sum([n_samples if left[node] != right[node] else 0 for node, n_samples in enumerate(samples)]) 466 | 467 | @_check_if_trained 468 | def get_top_branches(self, top_k=10): 469 | """ 470 | Returns list of top-k _branches of the best student, sorted by the number of samples the branch classifies. 471 | 472 | Parameters 473 | ---------- 474 | top_k: int, default=10 475 | Number of top-k branches, sorted by number of samples per branch, to return. 476 | 477 | Returns 478 | ------- 479 | top_branches: array-like of dict 480 | Dict of top-k branches from top student model. 481 | 482 | - dict: { "level": int, "path": array-like of dict, "class": int, "prob": float, "samples": int} 483 | """ 484 | if not self._branches: 485 | self._features, self._nodes, self._branches = get_dt_info(self._best_student) 486 | 487 | return sorted(self._branches, key=lambda p: p["samples"], reverse=True)[:top_k] 488 | 489 | @_check_if_trained 490 | def get_top_features(self, top_k=10): 491 | """ 492 | Get list of top _features of the best student, sorted by the number of samples the feature is used to classify. 493 | 494 | Parameters 495 | ---------- 496 | top_k: int, default=10 497 | Number of top-k features, sorted by number of samples per branch, to return. 498 | 499 | 500 | Returns 501 | ------- 502 | top_features: array-like of dict 503 | List of top-k features from top student model. 504 | 505 | - dict {"(int)" : {"count": int"samples": int}} 506 | """ 507 | if not self._features: 508 | self._features, self._nodes, self._branches = get_dt_info(self._best_student) 509 | 510 | return sorted(self._features.items(), key=lambda p: p[1]["samples"], reverse=True)[:top_k] 511 | 512 | @_check_if_trained 513 | def get_top_nodes(self, top_k=10): 514 | """ 515 | Returns list of top _nodes of the best student, sorted by the proportion of samples split by each node. 516 | 517 | The proportion of samples is calculated based on the impurity decrease equation is the following:: 518 | n_samples * abs(left_impurity - right_impurity) 519 | 520 | Parameters 521 | ---------- 522 | top_k: int, default=10 523 | Number of top-k nodes, sorted by number of samples per branch, to return. 524 | 525 | Returns 526 | ------- 527 | top_nodes: array-like of dict 528 | List of top-k nodes from top student model. 529 | 530 | - dict: {"idx": int, "level": int, "feature": int, "threshold": float, "samples": int, 531 | "values": tuple of int, "gini_split": tuple of float, "data_split": tuple of float, 532 | "data_split_by_class": array-like of tuple of float} 533 | """ 534 | if not self._nodes: 535 | self._features, self._nodes, self._branches = get_dt_info(self._best_student) 536 | 537 | return sorted( 538 | self._nodes, key=lambda p: p["samples"] * abs(p["gini_split"][0] - p["gini_split"][1]), reverse=True 539 | )[:top_k] 540 | 541 | @_check_if_trained 542 | def get_samples_by_level(self): 543 | """ 544 | Get number of samples by level of the best student. 545 | 546 | Returns 547 | ------- 548 | samples_by_level: dict of int 549 | Dict of samples by level. {"(int)": (int)} 550 | """ 551 | if not self._nodes: 552 | self._features, self._nodes, self._branches = get_dt_info(self._best_student) 553 | 554 | samples_by_level = list(np.zeros(self._best_student.get_depth() + 1)) 555 | for node in self._nodes: 556 | samples_by_level[node["level"]] += node["samples"] 557 | 558 | for node in self._branches: 559 | samples_by_level[node["level"]] += node["samples"] 560 | 561 | return samples_by_level 562 | 563 | @_check_if_trained 564 | def get_leaves_by_level(self): 565 | """ 566 | Returns number of leaves by level of the best student. 567 | 568 | Returns 569 | ------- 570 | leaves_by_level: dict of int 571 | Dict of leaves by level. {"(int)": (int)} 572 | """ 573 | if not self._branches: 574 | self._features, self._nodes, self._branches = get_dt_info(self._best_student) 575 | 576 | leaves_by_level = list(np.zeros(self._best_student.get_depth() + 1).astype(int)) 577 | for node in self._branches: 578 | leaves_by_level[node["level"]] += 1 579 | 580 | return leaves_by_level 581 | 582 | @_check_if_trained 583 | def prune(self, top_k=10, max_impurity=0.10): 584 | """ 585 | Prunes and returns the best student model explanation from the list of _students_by_iter. 586 | 587 | Parameters 588 | ---------- 589 | top_k: int, default=10 590 | Number of top-k branches, sorted by number of samples per branch, to return. 591 | 592 | max_impurity: float, default=0.10 593 | Maximum impurity allowed in a branch. Will prune anything below that impurity level. 594 | 595 | Returns 596 | ------- 597 | top_k_pruned_student: {DecisionTreeClassifier, DecisionTreeRegressor} 598 | Top-k pruned best fitted student model. 599 | """ 600 | return top_k_prune(self._best_student, top_k=top_k, max_impurity=max_impurity) 601 | 602 | 603 | class ClassificationTrustee(Trustee): 604 | """ 605 | Implements the Trust-oriented Decision Tree Extraction (Trustee) algorithm to train 606 | a student DecisionTreeClassifier based on observations from an Expert classification model. 607 | """ 608 | 609 | def __init__(self, expert, logger=None): 610 | """ 611 | Classification Trustee constructor 612 | 613 | Parameters 614 | ---------- 615 | expert: object 616 | The ML blackbox model to analyze. The expert model must have a `predict` method call implemented for 617 | Trustee to work properly, unless explicitly stated otherwise using the `predict_method_name` argument 618 | in the fit() method. 619 | logger: Logger object , default=None 620 | A logger object to log messages to. If none is given, the print() method will be used to log messages. 621 | """ 622 | super().__init__(expert, student_class=DecisionTreeClassifier, logger=logger) 623 | 624 | def _score(self, y_true, y_pred, average="macro"): 625 | """ 626 | Score function for student models. Compares the ground-truth predictions 627 | of a blackbox model with the predictions of a student model, using F1-score. 628 | 629 | Parameters 630 | ---------- 631 | y_true: array-like of shape (n_samples,) or (n_samples, n_outputs) 632 | The ground-truth target values (class labels in classification, real numbers in regression). 633 | 634 | y_pred: array-like of shape (n_samples,) or (n_samples, n_outputs) 635 | The predicted target values (class labels in classification, real numbers in regression). 636 | 637 | Returns 638 | ------- 639 | score: float 640 | Calculated F1-score between student model predictions and expert model ground-truth. 641 | """ 642 | return f1_score(y_true, y_pred, average=average) 643 | 644 | 645 | class RegressionTrustee(Trustee): 646 | """ 647 | Implements the Trust-oriented Decision Tree Extraction (Trustee) algorithm to train a 648 | student DecisionTreeRegressor based on observations from an Expert regression model. 649 | """ 650 | 651 | def __init__(self, expert, logger=None): 652 | """ 653 | Regression Trustee constructor 654 | 655 | Parameters 656 | ---------- 657 | expert: object 658 | The ML blackbox model to analyze. The expert model must have a `predict` method call implemented for 659 | Trustee to work properly, unless explicitly stated otherwise using the `predict_method_name` argument 660 | in the fit() method. 661 | logger: Logger object , default=None 662 | A logger object to log messages to. If none is given, the print() method will be used to log messages. 663 | """ 664 | super().__init__(expert=expert, student_class=DecisionTreeRegressor, logger=logger) 665 | 666 | def _score(self, y_true, y_pred): 667 | """ 668 | Score function for student models. Compares the ground-truth predictions 669 | of a blackbox model with the predictions of a student model, using R2-score. 670 | 671 | Parameters 672 | ---------- 673 | y_true: array-like of shape (n_samples,) or (n_samples, n_outputs) 674 | The ground-truth target values (class labels in classification, real numbers in regression). 675 | 676 | y_pred: array-like of shape (n_samples,) or (n_samples, n_outputs) 677 | The predicted target values (class labels in classification, real numbers in regression). 678 | 679 | Returns 680 | ------- 681 | score: float 682 | Calculated R2-score between student model predictions and expert model ground-truth. 683 | """ 684 | return r2_score(y_true, y_pred) 685 | -------------------------------------------------------------------------------- /trustee/report/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TrusteeML/trustee/71561fbcc96af8596fbd23879c3edcf0639e985b/trustee/report/__init__.py -------------------------------------------------------------------------------- /trustee/report/plot.py: -------------------------------------------------------------------------------- 1 | import os 2 | import re 3 | import math 4 | import numbers 5 | import numpy as np 6 | import pandas as pd 7 | 8 | from copy import deepcopy 9 | from pandas.api.types import is_numeric_dtype 10 | 11 | import matplotlib.pyplot as plt 12 | from matplotlib.ticker import PercentFormatter 13 | from sklearn.metrics import f1_score, r2_score 14 | 15 | from trustee.utils import plot 16 | from trustee.utils.tree import get_dt_info 17 | 18 | 19 | def plot_top_features(top_features, dt_sum_samples, dt_nodes, output_dir, feature_names=[]): 20 | """Uses top features information and plots CDF with it""" 21 | if not np.array(top_features).size or not np.array(dt_sum_samples).size or not np.array(dt_nodes).size: 22 | return 23 | 24 | features = [feature_names[feat] if feature_names else str(feat) for (feat, _) in top_features] 25 | count = [(values["count"] / dt_nodes) * 100 for (_, values) in top_features] 26 | count_sum = np.cumsum(count) 27 | data = [(values["samples"] / dt_sum_samples) * 100 for (_, values) in top_features] 28 | data_sum = np.cumsum(data) 29 | 30 | plot.plot_lines( 31 | features, 32 | [count_sum, data_sum], 33 | ylim=(0, 100), 34 | xlabel="Feature", 35 | ylabel="% of total", 36 | labels=["Nodes", "Samples"], 37 | path=f"{output_dir}/top_features_lines.pdf", 38 | ) 39 | plot.plot_bars( 40 | features, 41 | [count_sum, data_sum], 42 | ylim=(0, 100), 43 | xlabel="Feature", 44 | ylabel="% of total", 45 | labels=["Nodes", "Samples"], 46 | path=f"{output_dir}/top_features_bars.pdf", 47 | ) 48 | 49 | plot.plot_lines_and_bars( 50 | features, 51 | [count_sum, data_sum], 52 | [count, data], 53 | ylim=(0, 100), 54 | xlabel="Feature", 55 | ylabel="% of total", 56 | labels=["Nodes", "Samples"], 57 | path=f"{output_dir}/top_features_lines_bars.pdf", 58 | ) 59 | 60 | 61 | def plot_top_nodes(top_nodes, dt_samples_by_class, dt_samples, output_dir, feature_names=[], class_names=[]): 62 | """Uses top features information and plots CDF with it""" 63 | if not np.array(top_nodes).size or not np.array(dt_samples_by_class).size or not np.array(dt_samples).size: 64 | return 65 | 66 | plot.plot_stacked_bars_split( 67 | [ 68 | "{} <= {:.2f}".format( 69 | feature_names[node["feature"]] if feature_names else node["feature"], 70 | node["threshold"], 71 | ) 72 | for node in top_nodes 73 | ], 74 | [[(node["data_split"][0] / dt_samples) * 100 for node in top_nodes]], 75 | [[(node["data_split"][1] / dt_samples) * 100 for node in top_nodes]], 76 | y_placeholder=[100], 77 | ylim=(0, 100), 78 | xlabel="Node", 79 | ylabel="% of total samples", 80 | path=f"{output_dir}/top_nodes.pdf", 81 | ) 82 | 83 | plot.plot_stacked_bars_split( 84 | [ 85 | "{} <= {:.2f}".format( 86 | feature_names[node["feature"]] if feature_names else node["feature"], 87 | node["threshold"], 88 | ) 89 | for node in top_nodes 90 | ], 91 | [ 92 | [(node["data_split_by_class"][idx][0] / dt_samples) * 100 for node in top_nodes] 93 | for idx in range(len(dt_samples_by_class)) 94 | ], 95 | [ 96 | [(node["data_split_by_class"][idx][1] / dt_samples) * 100 for node in top_nodes] 97 | for idx in range(len(dt_samples_by_class)) 98 | ], 99 | y_placeholder=[(samples / dt_samples) * 100 for samples in dt_samples_by_class], 100 | ylim=(0, 100), 101 | xlabel="Node", 102 | ylabel="% of total samples", 103 | labels=class_names if class_names is not None else [], 104 | path=f"{output_dir}/top_nodes_by_class.pdf", 105 | ) 106 | 107 | 108 | def plot_top_branches( 109 | top_branches, 110 | dt_samples_by_class, 111 | dt_samples, 112 | output_dir, 113 | filename="top_branches", 114 | class_names=[], 115 | is_classify=True, 116 | ): 117 | """Uses top features information and plots CDF with it""" 118 | if not np.array(top_branches).size or not np.array(dt_samples_by_class).size or not np.array(dt_samples).size: 119 | return 120 | 121 | colors = [ 122 | "#d75d5b", 123 | "#524a47", 124 | "#8a4444", 125 | "#edeef0", 126 | "#c8c5c3", 127 | "#f5f0ed", 128 | "#a7c3cd", 129 | ] 130 | 131 | samples = [] 132 | colors_by_class = {} 133 | colors_by_samples = [] 134 | for branch in top_branches: 135 | class_label = class_names[branch["class"]] if class_names is not None else branch["class"] 136 | if class_label not in colors_by_class: 137 | colors_by_class[class_label] = ( 138 | colors.pop() if colors else "#%02x%02x%02x" % tuple(np.random.randint(256, size=3)) 139 | ) 140 | samples.append((branch["samples"] / dt_samples) * 100) 141 | colors_by_samples.append(colors_by_class[class_label]) 142 | 143 | plot.plot_lines( 144 | range(len(top_branches[:600])), 145 | [np.cumsum(samples[:600])], 146 | ylim=(0, 100), 147 | xlabel="Top-k Branches", 148 | ylabel="% of Samples", 149 | path=f"{output_dir}/{filename}.pdf", 150 | ) 151 | 152 | plot.plot_lines_and_bars( 153 | [f"Top {idx + 1}" for idx in range(len(top_branches))] if len(top_branches) < 20 else range(len(top_branches)), 154 | [np.cumsum(samples)], 155 | [samples], 156 | ylim=(0, 100), 157 | xlabel="Branches", 158 | ylabel="% of total samples", 159 | legend={"CDF": "#d75d5b", **colors_by_class}, 160 | colors_by_x=colors_by_samples, 161 | path=f"{output_dir}/{filename}_by_class.pdf", 162 | ) 163 | 164 | # TODO: This only works for classification problems, fix for refression in the future. 165 | if is_classify: 166 | plot.plot_stacked_bars( 167 | [f"Top {idx + 1}" for idx in range(len(top_branches))] 168 | if len(top_branches) < 20 169 | else range(len(top_branches)), 170 | [np.cumsum(samples)], 171 | y_placeholder=[100], 172 | ylim=(0, 100), 173 | xlabel="Branches", 174 | ylabel="% of total samples", 175 | path=f"{output_dir}/{filename}_bars.pdf", 176 | ) 177 | 178 | plot.plot_stacked_bars( 179 | [f"Top {idx + 1}" for idx in range(len(top_branches))] 180 | if len(top_branches) < 20 181 | else range(len(top_branches)), 182 | [ 183 | np.cumsum( 184 | [ 185 | ((branch["samples"] / dt_samples) * 100) if idx == branch["class"] else 0 186 | for branch in top_branches 187 | ] 188 | ) 189 | for idx, _ in enumerate(dt_samples_by_class) 190 | ], 191 | y_placeholder=[(samples / dt_samples) * 100 for samples in dt_samples_by_class], 192 | ylim=(0, 100), 193 | xlabel="Branches", 194 | ylabel="% of total samples", 195 | labels=class_names, 196 | path=f"{output_dir}/cum_{filename}_by_class.pdf", 197 | ) 198 | 199 | 200 | def plot_all_branches(top_branches, dt_samples_by_class, dt_samples, output_dir, class_names=[], is_classify=True): 201 | """Uses all features information and plots CDF with it""" 202 | plot_top_branches( 203 | top_branches, 204 | dt_samples_by_class, 205 | dt_samples, 206 | output_dir, 207 | filename="all_branches", 208 | class_names=class_names, 209 | is_classify=is_classify, 210 | ) 211 | 212 | 213 | def plot_samples_by_level(dt_samples_by_level, dt_nodes_by_level, dt_samples, output_dir): 214 | """Uses dt information to plot number of samples per level""" 215 | if not np.array(dt_nodes_by_level).size or not np.array(dt_nodes_by_level).size or not np.array(dt_samples).size: 216 | return 217 | 218 | samples = [] 219 | for idx, level_samples in enumerate(dt_samples_by_level): 220 | if idx < len(dt_samples_by_level) - 1: 221 | samples.append(((level_samples - dt_samples_by_level[idx + 1]) / dt_samples) * 100) 222 | else: 223 | samples.append((level_samples / dt_samples) * 100) 224 | 225 | plot.plot_lines_and_bars( 226 | [level for level, _ in enumerate(dt_samples_by_level)], 227 | [np.cumsum(samples)], 228 | [samples], 229 | ylim=(0, 100), 230 | xlabel="Level", 231 | ylabel="% of total samples", 232 | second_x_axis=dt_nodes_by_level, 233 | second_x_axis_label="Leaves at Level", 234 | labels=["Samples"], 235 | legend={"CDF": "#d75d5b", "Samples": "#c8c5c3"}, 236 | path=f"{output_dir}/samples_by_level.pdf", 237 | ) 238 | 239 | 240 | def plot_dts_fidelity_by_size(pruning_list, output_dir, filename="dts"): 241 | """Uses pruning information to plot fidelity vs size of decision trees""" 242 | if not np.array(pruning_list).size: 243 | return 244 | 245 | num_leaves = {} 246 | depth = {} 247 | fidelity = {} 248 | 249 | for pr in pruning_list: 250 | for i in pr["iter"]: 251 | if pr["type"] not in num_leaves: 252 | num_leaves[pr["type"]] = [] 253 | depth[pr["type"]] = [] 254 | fidelity[pr["type"]] = [] 255 | 256 | num_leaves[pr["type"]].append(i["dt"].get_n_leaves()) 257 | depth[pr["type"]].append(i["dt"].get_depth()) 258 | fidelity[pr["type"]].append(i["fidelity"]) 259 | 260 | plot.plot_lines( 261 | list(num_leaves.values()), 262 | list(fidelity.values()), 263 | ylim=(0, 1), 264 | xlim=(0, 50), 265 | xlabel="Number of Branches", 266 | ylabel="Fidelity", 267 | labels=list(num_leaves.keys()), 268 | path=f"{output_dir}/{filename}_fidelity_x_leaves.pdf", 269 | ) 270 | 271 | plot.plot_lines( 272 | list(depth.values()), 273 | list(fidelity.values()), 274 | ylim=(0, 1), 275 | xlabel="Depth", 276 | ylabel="Fidelity", 277 | labels=list(depth.keys()), 278 | path=f"{output_dir}/{filename}_fidelity_x_depth.pdf", 279 | ) 280 | 281 | 282 | def plot_stability( 283 | stability_iter, 284 | X_test, 285 | y_test, 286 | base_tree, 287 | base_tree_key, 288 | top_branches, 289 | output_dir, 290 | class_names=[], 291 | is_classify=True, 292 | ): 293 | """Uses stability information to plot the edit-distance between decision trees""" 294 | if not np.array(stability_iter).size: 295 | return 296 | 297 | agreement = [] 298 | agreement_by_class = {} 299 | nodes = {} 300 | features = {} 301 | features_by_it = {} 302 | total_nodes = 0 303 | number_of_splits = [] 304 | fidelity = [] 305 | base_y_pred = base_tree.predict(X_test.values) 306 | base_df = pd.DataFrame(deepcopy(X_test)) 307 | base_df["label"] = y_test 308 | grouped_df = base_df.groupby("label") if is_classify else [] 309 | 310 | for idx, it in enumerate(stability_iter): 311 | iter_tree = it[f"{base_tree_key}"] 312 | _, splits, _ = get_dt_info(iter_tree) 313 | total_nodes += len(splits) 314 | features_by_it[idx] = 0 315 | number_of_splits.append(len(splits)) 316 | 317 | for split in splits: 318 | split_str = f"{split['feature']}-{split['threshold']}" 319 | if split_str not in nodes: 320 | nodes[split_str] = 0 321 | nodes[split_str] += 1 322 | 323 | if split["feature"] not in features: 324 | features[split["feature"]] = {} 325 | 326 | if idx not in features[split["feature"]]: 327 | features[split["feature"]][idx] = 0 328 | features_by_it[idx] += 1 329 | features[split["feature"]][idx] += 1 330 | 331 | y_pred = iter_tree.predict(X_test.values) 332 | fidelity.append(it[f"{base_tree_key}_fidelity"]) 333 | agreement.append( 334 | f1_score(y_pred, base_y_pred, average="weighted") if is_classify else r2_score(y_pred, base_y_pred) 335 | ) 336 | 337 | for group, data in grouped_df: 338 | y_pred_class = iter_tree.predict(data.drop("label", axis=1).values) 339 | base_y_pred_class = base_tree.predict(data.drop("label", axis=1).values) 340 | if group not in agreement_by_class: 341 | agreement_by_class[group] = [] 342 | 343 | agreement_by_class[group].append(f1_score(y_pred_class, base_y_pred_class, average="weighted")) 344 | 345 | plot.plot_lines( 346 | range(len(number_of_splits)), 347 | [number_of_splits], 348 | xlabel="Iteration", 349 | ylabel="Number of Splits", 350 | path=f"{output_dir}/{base_tree_key}_num_nodes_stability.pdf", 351 | ) 352 | 353 | plot.plot_lines( 354 | range(len(features_by_it.keys())), 355 | [features_by_it.values()], 356 | xlabel="Iteration", 357 | ylabel="Stability", 358 | labels=["Features"], 359 | path=f"{output_dir}/{base_tree_key}_feature_stability.pdf", 360 | ) 361 | 362 | plot.plot_lines( 363 | range(len(stability_iter)), 364 | [agreement, fidelity], 365 | ylim=(0, 1), 366 | xlabel="Iteration", 367 | ylabel="Score", 368 | labels=["Agreement", "Fidelity"], 369 | path=f"{output_dir}/{base_tree_key}_stability.pdf", 370 | ) 371 | 372 | if is_classify: 373 | top_branch_agreement = {} 374 | for branch in top_branches[:5]: 375 | class_name = class_names[branch["class"]] if class_names is not None else branch["class"] 376 | class_id = class_name if class_name in agreement_by_class else branch["class"] 377 | top_branch_agreement[class_id] = agreement_by_class[class_id] 378 | 379 | plot.plot_lines( 380 | range(len(stability_iter)), 381 | [agreement for _, agreement in top_branch_agreement.items()], 382 | ylim=(0, 1), 383 | xlabel="Iteration", 384 | ylabel="Agreement (Score)", 385 | labels=[ 386 | class_names[group] if class_names is not None and not isinstance(group, str) else group 387 | for group, _ in top_branch_agreement.items() 388 | ], 389 | path=f"{output_dir}/{base_tree_key}_stability_by_class.pdf", 390 | size=(6, 4), 391 | ) 392 | 393 | 394 | def plot_stability_heatmap( 395 | stability_iter, 396 | X_test, 397 | y_test, 398 | tree_key, 399 | top_branches, 400 | output_dir, 401 | class_names=[], 402 | is_classify=True, 403 | ): 404 | """Uses stability information to plot the edit-distance between decision trees""" 405 | if not np.array(stability_iter).size: 406 | return 407 | 408 | heatmap_size = 30 409 | agreement = [] 410 | fidelity = [] 411 | mean_agreement = [] 412 | agreement_by_class = {} 413 | base_df = pd.DataFrame(deepcopy(X_test)) 414 | base_df["label"] = y_test 415 | grouped_df = base_df.groupby("label") if is_classify else [] 416 | 417 | for i, _ in enumerate(stability_iter): 418 | base_tree = stability_iter[i][f"{tree_key}"] 419 | fidelity.append(stability_iter[i][f"{tree_key}_fidelity"]) 420 | agreement.append([]) 421 | 422 | for j, _ in enumerate(stability_iter): 423 | iter_tree = stability_iter[j][f"{tree_key}"] 424 | 425 | iter_y_pred = iter_tree.predict(X_test.values) 426 | base_y_pred = base_tree.predict(X_test.values) 427 | 428 | agreement[i].append( 429 | f1_score(iter_y_pred, base_y_pred, average="weighted") 430 | if is_classify 431 | else r2_score(iter_y_pred, base_y_pred) 432 | ) 433 | 434 | for group, data in grouped_df: 435 | y_pred_class = iter_tree.predict(data.drop("label", axis=1).values) 436 | base_y_pred_class = base_tree.predict(data.drop("label", axis=1).values) 437 | if group not in agreement_by_class: 438 | agreement_by_class[group] = [] 439 | 440 | if i >= len(agreement_by_class[group]): 441 | agreement_by_class[group].append([]) 442 | 443 | agreement_by_class[group][i].append( 444 | f1_score(y_pred_class, base_y_pred_class, average="weighted") 445 | if is_classify 446 | else r2_score(y_pred_class, base_y_pred_class) 447 | ) 448 | mean_agreement.append(np.mean(agreement[i])) 449 | 450 | plot.plot_lines( 451 | range(len(stability_iter)), 452 | [mean_agreement, fidelity], 453 | ylim=(0, 1), 454 | xlim=(0, 50), 455 | xlabel="Iteration", 456 | ylabel="Score", 457 | labels=["Mean Agreement", "Fidelity"], 458 | path=f"{output_dir}/{tree_key}_mean_stability.pdf", 459 | ) 460 | 461 | plot.plot_heatmap( 462 | np.array([arr[:heatmap_size] for arr in agreement[:heatmap_size]]), 463 | labels=range(min(len(stability_iter), heatmap_size)), 464 | path=f"{output_dir}/{tree_key}_stability_heatmap.pdf", 465 | ) 466 | 467 | if is_classify: 468 | top_branch_agreement = {} 469 | for branch in top_branches[:5]: 470 | class_name = class_names[branch["class"]] if class_names is not None else branch["class"] 471 | class_id = class_name if class_name in agreement_by_class else branch["class"] 472 | top_branch_agreement[class_id] = agreement_by_class[class_id] 473 | 474 | for group, group_agreement in top_branch_agreement.items(): 475 | plot.plot_heatmap( 476 | np.array(group_agreement[:heatmap_size]), 477 | labels=range(min(len(stability_iter), heatmap_size)), 478 | path=f"{output_dir}/{tree_key}_{class_names[group] if class_names is not None and not isinstance(group, str) else group}_stability_heatmap.pdf", 479 | ) 480 | 481 | 482 | def plot_accuracy_by_feature_removed(whitebox_iter, output_dir, feature_names=[]): 483 | """Uses iterative analysis information to plot f1-score from the trained blackbox vs number of features removed""" 484 | if not np.array(whitebox_iter).size: 485 | return 486 | 487 | blackbox_scores = [i["score"] * 100 for i in whitebox_iter] 488 | fidelity = [i["fidelity"] * 100 for i in whitebox_iter] 489 | features = [feature_names[i["feature_removed"]] if feature_names else i["feature_removed"] for i in whitebox_iter] 490 | plot.plot_lines( 491 | features, 492 | [blackbox_scores, fidelity], 493 | ylim=(0, 100), 494 | xlabel="Features removed", 495 | ylabel="Metric (%)", 496 | labels=["Blackbox Score", "DT Fidelity"], 497 | path=f"{output_dir}/accuracy_by_feature_removed.pdf", 498 | ) 499 | 500 | 501 | def plot_distribution(X, y, top_branches, output_dir, aggregate=False, feature_names=[], class_names=[]): 502 | if isinstance(X, pd.DataFrame): 503 | X = X.values 504 | if isinstance(y, pd.Series): 505 | y = y.values 506 | 507 | """Plots the distribution of the data based on the top branches""" 508 | if not np.array(X).size or not np.array(y).size or not np.array(top_branches).size: 509 | return 510 | 511 | plots_output_dir = f"{output_dir}/dist" if not aggregate else f"{output_dir}/aggr_dist" 512 | if not os.path.exists(plots_output_dir): 513 | os.makedirs(plots_output_dir) 514 | 515 | colors = [ 516 | "#d75d5b", 517 | "#524a47", 518 | "#8a4444", 519 | "#edeef0", 520 | "#c8c5c3", 521 | "#f5f0ed", 522 | "#a7c3cd", 523 | ] 524 | 525 | df = pd.DataFrame(X, columns=feature_names if feature_names else None) 526 | if isinstance(df.columns[0], numbers.Number): 527 | df.columns = [str(i) for i in range(len(df.columns))] 528 | 529 | if aggregate: 530 | col_regex = "([\w_]+)_([0-9]+)" 531 | opt_prefixes = set({}) 532 | non_opt_prefixes = set({}) 533 | field_size = {} 534 | non_aggr_cols = df.columns # for plotting 535 | for col in df.columns: 536 | match_groups = re.findall(col_regex, col)[0] 537 | prefix = match_groups[0] 538 | bit = int(match_groups[1]) 539 | if "opt" in col: 540 | opt_prefixes.add(prefix) 541 | else: 542 | non_opt_prefixes.add(prefix) 543 | 544 | if prefix not in field_size: 545 | field_size[prefix] = bit 546 | 547 | if field_size[prefix] < bit: 548 | field_size[prefix] = bit 549 | 550 | # we need to treat option differently 551 | opt_df = df[[col for col in df.columns if "opt" in col]] 552 | non_opt_df = df[[col for col in df.columns if "opt" not in col]] 553 | 554 | def bin_to_int(num): 555 | try: 556 | return int(num, 2) 557 | except: 558 | return -1 559 | 560 | grouper = [next(p for p in non_opt_prefixes if p in c) for c in non_opt_df.columns] 561 | non_opt_df = non_opt_df.groupby(grouper, axis=1).apply( 562 | lambda x: x.astype(str).apply("".join, axis=1).apply(bin_to_int) 563 | ) 564 | df = pd.concat([non_opt_df, opt_df], axis=1) 565 | 566 | df["label"] = y 567 | if class_names is not None and is_numeric_dtype(df["label"]): 568 | df["label"] = df["label"].map(lambda x: class_names[int(x)]) 569 | 570 | num_classes = len(np.unique(y)) 571 | split_dfs = [x for _, x in df.groupby("label")] 572 | 573 | for idx, branch in enumerate(top_branches): 574 | branch_class = class_names[branch["class"]] if class_names is not None else str(branch["class"]) 575 | branch_output_dir = f"{plots_output_dir}/{idx}_branch_{branch_class}" 576 | 577 | if not os.path.exists(branch_output_dir): 578 | os.makedirs(branch_output_dir) 579 | 580 | filtered_dfs = [x.copy(deep=True) for _, x in df.groupby("label")] 581 | for rule_idx, (_, feat, op, thresh) in enumerate(branch["path"]): 582 | if aggregate: 583 | column = non_aggr_cols[int(feat)] if (isinstance(feat, numbers.Number) or feat.isdigit()) else feat 584 | if "opt" not in column: 585 | match_groups = re.findall(col_regex, column)[0] 586 | column = match_groups[0] 587 | bit = match_groups[1] 588 | shift = field_size[column] 589 | thresh = (1 << (shift - int(bit))) - (1 - int(thresh)) 590 | else: 591 | column = df.columns[int(feat)] if (isinstance(feat, numbers.Number) or feat.isdigit()) else feat 592 | 593 | plots_per_row = 5 594 | if num_classes > plots_per_row: 595 | n_rows = math.gcd(num_classes, plots_per_row) 596 | n_cols = num_classes if num_classes <= plots_per_row else int(num_classes / n_rows) 597 | else: 598 | n_rows = num_classes 599 | n_cols = 1 600 | 601 | fig, axes = plt.subplots(nrows=n_rows, ncols=n_cols) 602 | axes = axes.flatten() 603 | for df_idx, split_df in enumerate(split_dfs): 604 | df_class = split_df["label"].unique()[0] 605 | 606 | ax = axes[df_idx] 607 | ax.hist( 608 | split_df[column].values, 609 | histtype="bar", 610 | label="All" if df_idx == 0 else None, 611 | color=colors[0], 612 | ) 613 | ax.yaxis.set_major_formatter(PercentFormatter(xmax=split_df.shape[0])) 614 | ax.tick_params(axis="both", labelsize=6) 615 | ax.set_title(df_class, fontsize=8) 616 | 617 | tlt = fig.suptitle(f"{column} {op} {thresh:.3f}") 618 | lgd = fig.legend(loc="lower center", bbox_to_anchor=(0.5, -0.05), fancybox=True, ncol=5) 619 | plt.tight_layout() 620 | plt.savefig( 621 | f"{branch_output_dir}/{rule_idx}_{column.replace('/', '_')}_hist_all.pdf", 622 | bbox_extra_artists=(lgd, tlt), 623 | bbox_inches="tight", 624 | ) 625 | plt.close() 626 | 627 | fig, axes = plt.subplots(nrows=n_rows, ncols=n_cols) 628 | axes = axes.flatten() 629 | for df_idx, split_df in enumerate(split_dfs): 630 | branch_filter = f"filtered_dfs[{df_idx}]['{column}'] {op} {thresh}" 631 | filtered_dfs[df_idx] = filtered_dfs[df_idx][eval(branch_filter)] 632 | df_class = split_df["label"].unique()[0] 633 | 634 | ax = axes[df_idx] 635 | ax.hist( 636 | filtered_dfs[df_idx][column].values, 637 | histtype="bar", 638 | label=f"Branch ({branch_class.strip()})" if df_idx == 0 else None, 639 | color=colors[-1], 640 | ) 641 | ax.yaxis.set_major_formatter(PercentFormatter(xmax=split_df.shape[0])) 642 | ax.tick_params(axis="both", labelsize=6) 643 | ax.set_title(df_class, fontsize=8) 644 | 645 | tlt = fig.suptitle(f"{column} {op} {thresh:.3f}") 646 | lgd = fig.legend(loc="lower center", bbox_to_anchor=(0.5, -0.05), fancybox=True, ncol=5) 647 | plt.tight_layout() 648 | plt.savefig( 649 | f"{branch_output_dir}/{rule_idx}_{column.replace('/', '_')}_hist_branch.pdf", 650 | bbox_extra_artists=(lgd, tlt), 651 | bbox_inches="tight", 652 | ) 653 | plt.close() 654 | 655 | fig, axes = plt.subplots(nrows=n_rows, ncols=n_cols) 656 | axes = axes.flatten() 657 | for df_idx, split_df in enumerate(split_dfs): 658 | df_class = split_df["label"].unique()[0] 659 | branch_filter = f"filtered_dfs[{df_idx}]['{column}'] {op} {thresh}" 660 | filtered_dfs[df_idx] = filtered_dfs[df_idx][eval(branch_filter)] 661 | 662 | ax = axes[df_idx] 663 | ax.hist( 664 | split_df[column].values, 665 | histtype="bar", 666 | label="All" if df_idx == 0 else None, 667 | color=colors[0], 668 | ) 669 | ax.hist( 670 | filtered_dfs[df_idx][column].values, 671 | histtype="bar", 672 | label=f"Branch ({branch_class.strip()})" if df_idx == 0 else None, 673 | color=colors[-1], 674 | ) 675 | ax.yaxis.set_major_formatter(PercentFormatter(xmax=split_df.shape[0])) 676 | ax.tick_params(axis="both", labelsize=6) 677 | ax.set_title(df_class, fontsize=8) 678 | 679 | tlt = fig.suptitle(f"{column} {op} {thresh:.3f}") 680 | lgd = fig.legend(loc="lower center", bbox_to_anchor=(0.5, -0.05), fancybox=True, ncol=5) 681 | plt.tight_layout() 682 | plt.savefig( 683 | f"{branch_output_dir}/{rule_idx}_{column.replace('/', '_')}_hist.pdf", 684 | bbox_extra_artists=(lgd, tlt), 685 | bbox_inches="tight", 686 | ) 687 | plt.close() 688 | -------------------------------------------------------------------------------- /trustee/utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TrusteeML/trustee/71561fbcc96af8596fbd23879c3edcf0639e985b/trustee/utils/__init__.py -------------------------------------------------------------------------------- /trustee/utils/const.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | 3 | from . import rootpath 4 | from trustee.enums.feature_type import FeatureType 5 | 6 | WINE_DATASET_META = { 7 | "name": "wine", 8 | "path": f"{rootpath.detect()}/res/dataset/wine.csv", 9 | "has_header": True, 10 | "delimiter": ";", 11 | "fields": [ 12 | ("fixed acidity", FeatureType.NUMERICAL, None, False), 13 | ("volatile acidity", FeatureType.NUMERICAL, None, False), 14 | ("citric acid", FeatureType.NUMERICAL, None, False), 15 | ("residual sugar", FeatureType.NUMERICAL, None, False), 16 | ("chlorides", FeatureType.NUMERICAL, None, False), 17 | ("free sulfur dioxide", FeatureType.NUMERICAL, None, False), 18 | ("total sulfur dioxide", FeatureType.NUMERICAL, None, False), 19 | ("density", FeatureType.NUMERICAL, None, False), 20 | ("pH", FeatureType.NUMERICAL, None, False), 21 | ("sulphates", FeatureType.NUMERICAL, None, False), 22 | ("alcohol", FeatureType.NUMERICAL, None, False), 23 | ("quality", FeatureType.NUMERICAL, None, True), 24 | ], 25 | "type": "regression", 26 | "url": "https://archive.ics.uci.edu/ml/datasets/wine", 27 | } 28 | 29 | DIABETES_DATASET_META = { 30 | "name": "diabetes", 31 | "path": f"{rootpath.detect()}/res/dataset/diabetes/diabetic_data.csv", 32 | "has_header": True, 33 | "delimiter": ",", 34 | "fields": [ 35 | ("encounter_id", FeatureType.IDENTIFIER, None, False), 36 | ("patient_nbr", FeatureType.IDENTIFIER, None, False), 37 | ("race", FeatureType.CATEGORICAL, None, False), 38 | ("gender", FeatureType.CATEGORICAL, None, False), 39 | ("age", FeatureType.CATEGORICAL, None, False), 40 | ("weight", FeatureType.CATEGORICAL, None, False), 41 | ("admission_type_id", FeatureType.CATEGORICAL, None, False), 42 | ("discharge_disposition_id", FeatureType.CATEGORICAL, None, False), 43 | ("admission_source_id", FeatureType.CATEGORICAL, None, False), 44 | ("time_in_hospital", FeatureType.NUMERICAL, None, False), 45 | ("payer_code", FeatureType.CATEGORICAL, None, False), 46 | ("medical_specialty", FeatureType.CATEGORICAL, None, False), 47 | ("num_lab_procedures", FeatureType.NUMERICAL, None, False), 48 | ("num_procedures", FeatureType.NUMERICAL, None, False), 49 | ("num_medications", FeatureType.NUMERICAL, None, False), 50 | ("number_outpatient", FeatureType.NUMERICAL, None, False), 51 | ("number_emergency", FeatureType.NUMERICAL, None, False), 52 | ("number_inpatient", FeatureType.NUMERICAL, None, False), 53 | ("diag_1", FeatureType.CATEGORICAL, None, False), 54 | ("diag_2", FeatureType.CATEGORICAL, None, False), 55 | ("diag_3", FeatureType.CATEGORICAL, None, False), 56 | ("number_diagnoses", FeatureType.NUMERICAL, None, False), 57 | ("max_glu_serum", FeatureType.CATEGORICAL, None, False), 58 | ("A1Cresult", FeatureType.CATEGORICAL, None, False), 59 | ("metformin", FeatureType.CATEGORICAL, None, False), 60 | ("repaglinide", FeatureType.CATEGORICAL, None, False), 61 | ("nateglinide", FeatureType.CATEGORICAL, None, False), 62 | ("chlorpropamide", FeatureType.CATEGORICAL, None, False), 63 | ("glimepiride", FeatureType.CATEGORICAL, None, False), 64 | ("acetohexamide", FeatureType.CATEGORICAL, None, False), 65 | ("glipizide", FeatureType.CATEGORICAL, None, False), 66 | ("glyburide", FeatureType.CATEGORICAL, None, False), 67 | ("tolbutamide", FeatureType.CATEGORICAL, None, False), 68 | ("pioglitazone", FeatureType.CATEGORICAL, None, False), 69 | ("rosiglitazone", FeatureType.CATEGORICAL, None, False), 70 | ("acarbose", FeatureType.CATEGORICAL, None, False), 71 | ("miglitol", FeatureType.CATEGORICAL, None, False), 72 | ("troglitazone", FeatureType.CATEGORICAL, None, False), 73 | ("tolazamide", FeatureType.CATEGORICAL, None, False), 74 | ("examide", FeatureType.CATEGORICAL, None, False), 75 | ("citoglipton", FeatureType.CATEGORICAL, None, False), 76 | ("insulin", FeatureType.CATEGORICAL, None, False), 77 | ("glyburide-metformin", FeatureType.CATEGORICAL, None, False), 78 | ("glipizide-metformin", FeatureType.CATEGORICAL, None, False), 79 | ("glimepiride-pioglitazone", FeatureType.CATEGORICAL, None, False), 80 | ("metformin-rosiglitazone", FeatureType.CATEGORICAL, None, False), 81 | ("metformin-pioglitazone", FeatureType.CATEGORICAL, None, False), 82 | ("change", FeatureType.CATEGORICAL, None, False), 83 | ("diabetesMed", FeatureType.CATEGORICAL, None, False), 84 | ("readmitted", FeatureType.CATEGORICAL, None, True), 85 | ], 86 | "type": "classification", 87 | "url": "https://www.kaggle.com/brandao/diabetes?select=diabetic_data.csv", 88 | } 89 | 90 | IOT_DATASET_META = { 91 | "name": "iot", 92 | # "path": "{}/res/dataset/iot/csv_files/16-09-23-labeled.csv".format(rootpath.detect()), 93 | "path": f"{rootpath.detect()}/res/dataset/iot/csv_files/", 94 | "is_dir": True, 95 | "has_header": False, 96 | "fields": [ 97 | ("Frame Length", FeatureType.NUMERICAL, None, False), 98 | ("Ethernet Type", FeatureType.NUMERICAL, None, False), 99 | ("IP Protocol", FeatureType.CATEGORICAL, None, False), 100 | ("IPv4 Flags", FeatureType.CATEGORICAL, None, False), 101 | ("IPv6 Next Header", FeatureType.CATEGORICAL, None, False), 102 | ("IPv6 Option", FeatureType.CATEGORICAL, None, False), 103 | ("TCP Src Port", FeatureType.NUMERICAL, None, False), 104 | ("TCP Dst Port", FeatureType.NUMERICAL, None, False), 105 | ("TCP Flags", FeatureType.CATEGORICAL, None, False), 106 | ("UDP Src Port", FeatureType.NUMERICAL, None, False), 107 | ("UDP Dst Port", FeatureType.NUMERICAL, None, False), 108 | ("IoT Device Type", FeatureType.CATEGORICAL, None, True), 109 | ], 110 | "classes": ["Smart Static", "Sensor", "Audio", "Video", "Other"], 111 | "converters": { 112 | 1: lambda x: int(x, 16) if x else None, 113 | 3: lambda x: int(x, 16) if x else None, 114 | 8: lambda x: int(x, 16) if x else None, 115 | }, 116 | "type": "classification", 117 | "categories": { 118 | "IP Protocol": [-1, 0, 1, 2, 6, 17, 145, 242], 119 | "IPv4 Flags": [ 120 | -1, 121 | 0, 122 | 185, 123 | 925, 124 | 8192, 125 | 8377, 126 | 8562, 127 | 8747, 128 | 8932, 129 | 16384, 130 | 48299, 131 | 60692, 132 | ], 133 | "IPv6 Next Header": [-1, 0, 6, 17, 44, 58], 134 | "IPv6 Option": [-1, 1], 135 | "TCP Flags": [ 136 | -1, 137 | 1, 138 | 2, 139 | 4, 140 | 16, 141 | 17, 142 | 18, 143 | 20, 144 | 24, 145 | 25, 146 | 28, 147 | 47, 148 | 49, 149 | 56, 150 | 82, 151 | 144, 152 | 152, 153 | 153, 154 | 168, 155 | 194, 156 | 210, 157 | 1041, 158 | 2050, 159 | 2051, 160 | 2513, 161 | 3345, 162 | 3610, 163 | ], 164 | }, 165 | "url": "https://iotanalytics.unsw.edu.au/iottraces.html", 166 | } 167 | 168 | BOSTON_DATASET_META = { 169 | "name": "boston", 170 | "path": f"{rootpath.detect()}/res/dataset/boston.csv", 171 | "has_header": True, 172 | "fields": [ 173 | ("CRIM", FeatureType.NUMERICAL, None, False), 174 | ("ZN", FeatureType.NUMERICAL, None, False), 175 | ("INDUS", FeatureType.NUMERICAL, None, False), 176 | ("CHAS", FeatureType.CATEGORICAL, None, False), 177 | ("NOX", FeatureType.NUMERICAL, None, False), 178 | ("RM", FeatureType.NUMERICAL, None, False), 179 | ("AGE", FeatureType.NUMERICAL, None, False), 180 | ("DIS", FeatureType.NUMERICAL, None, False), 181 | ("RAD", FeatureType.NUMERICAL, None, False), 182 | ("TAX", FeatureType.NUMERICAL, None, False), 183 | ("PTRATIO", FeatureType.NUMERICAL, None, False), 184 | ("B", FeatureType.NUMERICAL, None, False), 185 | ("LSTAT", FeatureType.NUMERICAL, None, False), 186 | ("MEDV", FeatureType.CATEGORICAL, None, True), 187 | ], 188 | "type": "regression", 189 | "url": "https://www.cs.toronto.edu/~delve/data/boston/bostonDetail.html", 190 | } 191 | 192 | 193 | def cic_ids_2017_label_converter(label): 194 | value = -1 195 | labels = { 196 | "BENIGN": 0, 197 | "Bot": 1, 198 | "DDoS": 2, 199 | "DoS GoldenEye": 3, 200 | "DoS Hulk": 4, 201 | "DoS Slowhttptest": 5, 202 | "DoS slowloris": 6, 203 | "FTP-Patator": 7, 204 | "Heartbleed": 8, 205 | "Infiltration": 9, 206 | "PortScan": 10, 207 | "SSH-Patator": 11, 208 | "Web Attack Brute Force": 12, 209 | "Web Attack Sql Injection": 13, 210 | "Web Attack XSS": 14, 211 | } 212 | 213 | try: 214 | value = labels[label.strip()] 215 | except Exception as err: 216 | print("Exception", err, label) 217 | 218 | return np.uint8(value) 219 | 220 | 221 | CIC_IDS_2017_DATASET_META = { 222 | "name": "cic_ids_2017", 223 | "path": f"{rootpath.detect()}/res/dataset/CIC-IDS-2017/MachineLearningCVE/", 224 | "is_dir": True, 225 | "oversampled_path": f"{rootpath.detect()}/res/dataset/CIC-IDS-2017/MachineLearningCVE_OverSampled.csv.zip", 226 | "undersampled_path": f"{rootpath.detect()}/res/dataset/CIC-IDS-2017/MachineLearningCVE_UnderSampled.csv", 227 | "has_header": True, 228 | "fields": [ 229 | ("Destination Port", FeatureType.NUMERICAL, "uint32", False), 230 | ("Flow Duration", FeatureType.NUMERICAL, "uint32", False), 231 | ("Total Fwd Packets", FeatureType.NUMERICAL, "uint32", False), 232 | ("Total Backward Packets", FeatureType.NUMERICAL, "uint32", False), 233 | ("Total Length of Fwd Packets", FeatureType.NUMERICAL, "uint32", False), 234 | ("Total Length of Bwd Packets", FeatureType.NUMERICAL, "uint32", False), 235 | ("Fwd Packet Length Max", FeatureType.NUMERICAL, "uint16", False), 236 | ("Fwd Packet Length Min", FeatureType.NUMERICAL, "uint16", False), 237 | ("Fwd Packet Length Mean", FeatureType.NUMERICAL, "float16", False), 238 | ("Fwd Packet Length Std", FeatureType.NUMERICAL, "float16", False), 239 | ("Bwd Packet Length Max", FeatureType.NUMERICAL, "uint16", False), 240 | ("Bwd Packet Length Min", FeatureType.NUMERICAL, "uint16", False), 241 | ("Bwd Packet Length Mean", FeatureType.NUMERICAL, "float16", False), 242 | ("Bwd Packet Length Std", FeatureType.NUMERICAL, "float16", False), 243 | ("Flow Bytes/s", FeatureType.NUMERICAL, "float32", False), 244 | ("Flow Packets/s", FeatureType.NUMERICAL, "float32", False), 245 | ("Flow IAT Mean", FeatureType.NUMERICAL, "float32", False), 246 | ("Flow IAT Std", FeatureType.NUMERICAL, "float32", False), 247 | ("Flow IAT Max", FeatureType.NUMERICAL, "uint32", False), 248 | ("Flow IAT Min", FeatureType.NUMERICAL, "uint32", False), 249 | ("Fwd IAT Total", FeatureType.NUMERICAL, "uint32", False), 250 | ("Fwd IAT Mean", FeatureType.NUMERICAL, "float32", False), 251 | ("Fwd IAT Std", FeatureType.NUMERICAL, "float32", False), 252 | ("Fwd IAT Max", FeatureType.NUMERICAL, "uint32", False), 253 | ("Fwd IAT Min", FeatureType.NUMERICAL, "uint32", False), 254 | ("Bwd IAT Total", FeatureType.NUMERICAL, "uint32", False), 255 | ("Bwd IAT Mean", FeatureType.NUMERICAL, "float32", False), 256 | ("Bwd IAT Std", FeatureType.NUMERICAL, "float32", False), 257 | ("Bwd IAT Max", FeatureType.NUMERICAL, "uint32", False), 258 | ("Bwd IAT Min", FeatureType.NUMERICAL, "uint32", False), 259 | ("Fwd PSH Flags", FeatureType.CATEGORICAL, "uint8", False), 260 | ("Bwd PSH Flags", FeatureType.IDENTIFIER, "uint8", False), 261 | ("Fwd URG Flags", FeatureType.CATEGORICAL, "uint8", False), 262 | ("Bwd URG Flags", FeatureType.IDENTIFIER, "uint8", False), 263 | ( 264 | "Fwd Header Length 2", 265 | FeatureType.IDENTIFIER, 266 | "uint32", 267 | False, 268 | ), # duplicate column, so ignore it 269 | ("Bwd Header Length", FeatureType.NUMERICAL, "uint32", False), 270 | ("Fwd Packets/s", FeatureType.NUMERICAL, "float16", False), 271 | ("Bwd Packets/s", FeatureType.NUMERICAL, "float16", False), 272 | ("Min Packet Length", FeatureType.NUMERICAL, "float16", False), 273 | ("Max Packet Length", FeatureType.NUMERICAL, "float16", False), 274 | ("Packet Length Mean", FeatureType.NUMERICAL, "float16", False), 275 | ("Packet Length Std", FeatureType.NUMERICAL, "float16", False), 276 | ("Packet Length Variance", FeatureType.NUMERICAL, "float32", False), 277 | ("FIN Flag Count", FeatureType.NUMERICAL, "uint8", False), 278 | ("SYN Flag Count", FeatureType.NUMERICAL, "uint8", False), 279 | ("RST Flag Count", FeatureType.NUMERICAL, "uint8", False), 280 | ("PSH Flag Count", FeatureType.NUMERICAL, "uint8", False), 281 | ("ACK Flag Count", FeatureType.NUMERICAL, "uint8", False), 282 | ("URG Flag Count", FeatureType.NUMERICAL, "uint8", False), 283 | ("CWE Flag Count", FeatureType.NUMERICAL, "uint8", False), 284 | ("ECE Flag Count", FeatureType.NUMERICAL, "uint8", False), 285 | ("Down/Up Ratio", FeatureType.NUMERICAL, "float16", False), 286 | ("Average Packet Size", FeatureType.NUMERICAL, "float16", False), 287 | ("Avg Fwd Segment Size", FeatureType.NUMERICAL, "float16", False), 288 | ("Avg Bwd Segment Size", FeatureType.NUMERICAL, "float16", False), 289 | ("Fwd Header Length", FeatureType.NUMERICAL, "uint32", False), 290 | ("Fwd Avg Bytes/Bulk", FeatureType.NUMERICAL, "uint8", False), 291 | ("Fwd Avg Packets/Bulk", FeatureType.NUMERICAL, "uint8", False), 292 | ("Fwd Avg Bulk Rate", FeatureType.NUMERICAL, "uint8", False), 293 | ("Bwd Avg Bytes/Bulk", FeatureType.NUMERICAL, "uint8", False), 294 | ("Bwd Avg Packets/Bulk", FeatureType.NUMERICAL, "uint8", False), 295 | ("Bwd Avg Bulk Rate", FeatureType.NUMERICAL, "uint8", False), 296 | ("Subflow Fwd Packets", FeatureType.NUMERICAL, "uint32", False), 297 | ("Subflow Fwd Bytes", FeatureType.NUMERICAL, "uint32", False), 298 | ("Subflow Bwd Packets", FeatureType.NUMERICAL, "uint32", False), 299 | ("Subflow Bwd Bytes", FeatureType.NUMERICAL, "uint32", False), 300 | ("Init_Win_bytes_forward", FeatureType.NUMERICAL, "uint32", False), 301 | ("Init_Win_bytes_backward", FeatureType.NUMERICAL, "uint32", False), 302 | ("act_data_pkt_fwd", FeatureType.NUMERICAL, "uint16", False), 303 | ("min_seg_size_forward", FeatureType.NUMERICAL, "uint16", False), 304 | ("Active Mean", FeatureType.NUMERICAL, "float32", False), 305 | ("Active Std", FeatureType.NUMERICAL, "float32", False), 306 | ("Active Max", FeatureType.NUMERICAL, "uint32", False), 307 | ("Active Min", FeatureType.NUMERICAL, "uint32", False), 308 | ("Idle Mean", FeatureType.NUMERICAL, "float32", False), 309 | ("Idle Std", FeatureType.NUMERICAL, "float32", False), 310 | ("Idle Max", FeatureType.NUMERICAL, "uint32", False), 311 | ("Idle Min", FeatureType.NUMERICAL, "uint32", False), 312 | ("Label", FeatureType.CATEGORICAL, "uint8", True), 313 | ], 314 | "categories": { 315 | "Fwd PSH Flags": [np.uint8(0), np.uint8(1)], 316 | # "Bwd PSH Flags": [np.uint8(0)], 317 | "Fwd URG Flags": [np.uint8(0), np.uint8(1)], 318 | # "Bwd URG Flags": [np.uint8(0)], 319 | }, 320 | "classes": [ 321 | "BENIGN", 322 | "Bot", 323 | "DDoS", 324 | "DoS GoldenEye", 325 | "DoS Hulk", 326 | "DoS Slowhttptest", 327 | "DoS slowloris", 328 | "FTP-Patator", 329 | "Heartbleed", 330 | "Infiltration", 331 | "PortScan", 332 | "SSH-Patator", 333 | "Web Attack Brute Force", 334 | "Web Attack Sql Injection", 335 | "Web Attack XSS", 336 | ], 337 | "converters": {"Label": lambda x: cic_ids_2017_label_converter(x)}, 338 | "type": "classification", 339 | "url": "https://www.unb.ca/cic/datasets/ids-2017.html", 340 | } 341 | -------------------------------------------------------------------------------- /trustee/utils/dataset.py: -------------------------------------------------------------------------------- 1 | import io 2 | import glob 3 | import numpy as np 4 | import pandas as pd 5 | 6 | from pandas.api.types import CategoricalDtype 7 | 8 | from trustee.enums.feature_type import FeatureType 9 | 10 | 11 | def convert_to_df(data): 12 | """ 13 | Converts data to a pandas DataFrame. 14 | 15 | Args: 16 | - data 17 | Data to be converted 18 | 19 | Returns: pd.DataFrame 20 | """ 21 | if isinstance(data, pd.DataFrame): 22 | return data 23 | elif isinstance(data, pd.Series): 24 | return pd.DataFrame(data) 25 | elif isinstance(data, np.ndarray): 26 | return pd.DataFrame(data) 27 | else: 28 | raise ValueError("Data must be a pandas dataframe or numpy array") 29 | 30 | 31 | def convert_to_series(data): 32 | """ 33 | Converts data to a pandas Series. 34 | 35 | Args: 36 | - data 37 | Data to be converted 38 | 39 | Returns: pd.Series 40 | """ 41 | if isinstance(data, pd.Series): 42 | return data 43 | elif isinstance(data, pd.DataFrame): 44 | return data.iloc[:, 0] 45 | elif isinstance(data, np.ndarray): 46 | return pd.Series(data.flatten()) 47 | else: 48 | raise ValueError("Data must be a pandas dataframe or numpy array") 49 | 50 | 51 | def read(path_or_buffer, metadata={}, verbose=False, logger=None, as_df=False): 52 | """ 53 | Reads dataset from a CSV and returns it as dataframe 54 | 55 | Args: 56 | - path 57 | Path to read dataset from 58 | Returns: 59 | - X 60 | Features from dataset as an np.array or dataframe (if as_df == True) 61 | - y 62 | Target variable from dataset as an np.array or dataframe (if as_df == True) 63 | - feature_names 64 | Feature names, if any were extracted from csv or metadata 65 | """ 66 | 67 | log = logger.log if logger else print 68 | 69 | X = [] 70 | y = [] 71 | 72 | idx_offset = 0 # cant use enumerate because of skipping ID features 73 | names = [] 74 | dtypes = {} 75 | numerical = [] 76 | categorical = [] 77 | dummies = [] 78 | use_cols = [] 79 | result = [] 80 | 81 | for idx, (name, type, dtype, is_result) in enumerate(metadata["fields"]): 82 | if type != FeatureType.IDENTIFIER: 83 | use_cols.append(idx) 84 | 85 | if dtype: 86 | dtypes[name] = dtype 87 | 88 | if name: 89 | names.append(name) 90 | 91 | if type == FeatureType.NUMERICAL and not is_result: 92 | numerical.append(idx - idx_offset) 93 | 94 | if type == FeatureType.CATEGORICAL and not is_result: 95 | dummies.append(idx - idx_offset) 96 | 97 | if is_result: 98 | result.append(idx - idx_offset) 99 | else: 100 | idx_offset += 1 101 | 102 | if verbose: 103 | log(10 * "=", "Metadata start.", 10 * "=") 104 | log("Names:", names) 105 | log("Dummies:", dummies) 106 | log("Use cols:", use_cols) 107 | log("Result variable:", result) 108 | log(10 * "=", "Metadata end.", 10 * "=") 109 | log("") # skip line 110 | 111 | if "is_dir" in metadata and metadata["is_dir"] and not isinstance(path_or_buffer, io.TextIOBase): 112 | df_list = [] 113 | 114 | datasets = glob.glob(path_or_buffer + "/*.csv") 115 | if not datasets: 116 | datasets = glob.glob(path_or_buffer + "/*.zip") 117 | 118 | for dataset_path in datasets: 119 | df = pd.read_csv( 120 | dataset_path, 121 | delimiter=metadata["delimiter"] if "delimiter" in metadata else ",", 122 | header=0 if "has_header" in metadata and metadata["has_header"] else None, 123 | names=names if names else None, 124 | dtype=dtypes if dtypes else None, 125 | usecols=use_cols, 126 | converters=metadata["converters"] if "converters" in metadata else None, 127 | ).fillna(-1) 128 | df_list.append(df) 129 | df = pd.concat(df_list, axis=0, ignore_index=True) 130 | else: 131 | df = pd.read_csv( 132 | path_or_buffer, 133 | delimiter=metadata["delimiter"] if "delimiter" in metadata else ",", 134 | header=0 if "has_header" in metadata and metadata["has_header"] else None, 135 | names=names if names else None, 136 | dtype=dtypes if dtypes else None, 137 | usecols=use_cols, 138 | converters=metadata["converters"] if "converters" in metadata else None, 139 | ).fillna(-1) 140 | 141 | names = df.columns # guaranteing names will be filled, even if read from csv 142 | if verbose: 143 | log("Pandas read_csv complete.") 144 | 145 | if "categories" in metadata: 146 | for (column, categories) in metadata["categories"].items(): 147 | category = CategoricalDtype(categories=categories, ordered=True) 148 | df[column] = df[column].astype(category) 149 | 150 | if verbose: 151 | log("CSV dataset read:") 152 | log(df) 153 | log(df.shape) 154 | log("Any NAN?", df.isnull().sum().sum()) 155 | total_usage_b = df.memory_usage(deep=True).sum() 156 | total_usage_mb = total_usage_b / 1024**2 157 | mean_usage_b = df.memory_usage(deep=True).mean() 158 | mean_usage_mb = mean_usage_b / 1024**2 159 | log(f"Total memory usage: {total_usage_mb:03.2f} MB") 160 | log(f"Average memory usage: {mean_usage_mb:03.2f} MB") 161 | 162 | # if no result varibles is passed in metadata, we assume the last column as the result 163 | if not result: 164 | result = [len(df.columns) - 1] 165 | 166 | y = df[names[result]].copy() 167 | 168 | X = df.drop(columns=names[result], axis=1) 169 | # resulting dataset corresponds to feature variables only, so encode it if necessary 170 | if dummies: 171 | dummy_cols = [names[i] for i in dummies] 172 | categorical = [[] for _ in dummy_cols] 173 | X = pd.get_dummies(X, columns=dummy_cols) 174 | for i, _ in enumerate(X.columns): 175 | for j, _ in enumerate(dummy_cols): 176 | cat_feat = dummy_cols[j] 177 | if str(X.columns[i]).startswith(f"{str(cat_feat)}_"): 178 | categorical[j].append(i) 179 | 180 | if verbose: 181 | log("Features Shape:", X.shape) 182 | column_names = "".join(f"{str(i)}: {str(col)}\n" for (i, col) in zip(list(range(len(X.columns))), X.columns)) 183 | log(f"Column names:\n{column_names}") 184 | log("Targets shape:", y.shape, y.columns) 185 | 186 | X = X.replace([np.inf, -np.inf], np.nan).fillna(-1) 187 | 188 | return ( 189 | X if as_df else np.nan_to_num(X.to_numpy()), 190 | y if as_df else y.to_numpy(), 191 | X.columns, 192 | numerical, 193 | categorical, 194 | ) 195 | -------------------------------------------------------------------------------- /trustee/utils/log.py: -------------------------------------------------------------------------------- 1 | import logging 2 | 3 | from . import rootpath 4 | 5 | 6 | class Logger(logging.getLoggerClass()): 7 | """Initialze log with output to given path""" 8 | 9 | def __init__( 10 | self, 11 | path=f"{rootpath.detect()}/res/log/output.log", 12 | level=logging.DEBUG, 13 | ): 14 | # Create handlers 15 | super().__init__(__name__) 16 | self.setLevel(level) 17 | 18 | stream_handler = logging.StreamHandler() 19 | file_handler = logging.FileHandler(path) 20 | stream_handler.setLevel(level) 21 | file_handler.setLevel(level) 22 | 23 | format = logging.Formatter("%(asctime)s - %(levelname)s - %(message)s") 24 | stream_handler.setFormatter(format) 25 | file_handler.setFormatter(format) 26 | 27 | # Add handlers to the log 28 | self.addHandler(stream_handler) 29 | self.addHandler(file_handler) 30 | 31 | def log(self, *args, level=logging.INFO): 32 | """Logs message with log with given level""" 33 | super().log(level, " ".join([str(arg) for arg in args])) 34 | -------------------------------------------------------------------------------- /trustee/utils/persist.py: -------------------------------------------------------------------------------- 1 | from copy import Error 2 | import os 3 | import zipfile 4 | import traceback 5 | 6 | from joblib import dump, load 7 | 8 | 9 | def load_model(path): 10 | """Loads model weights, if they exist""" 11 | file_path = path 12 | if zipfile.is_zipfile(path): 13 | file_dir = os.path.dirname(path) 14 | file_path = path.replace(".zip", "") 15 | zip_ref = zipfile.ZipFile(path) # create zipfile object 16 | zip_ref.extractall(file_dir) # extract file to dir 17 | zip_ref.close() # close file 18 | 19 | if file_path and os.path.isfile(file_path): 20 | try: 21 | model = load(file_path) 22 | if zipfile.is_zipfile(path): 23 | os.remove(file_path) # delete unzipped file 24 | 25 | return model 26 | except Error: 27 | traceback.print_exc() 28 | return None 29 | return None 30 | 31 | 32 | def save_model(model, path): 33 | """Saves model weights""" 34 | return dump(model, path) 35 | -------------------------------------------------------------------------------- /trustee/utils/plot.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import matplotlib 3 | import matplotlib.pyplot as plt 4 | import matplotlib.patches as mpatches 5 | 6 | from matplotlib import rcParams 7 | 8 | FONT_NAME = "Roboto" 9 | FONT_WEIGHT = "light" 10 | 11 | rcParams["font.family"] = "serif" 12 | rcParams["font.serif"] = [FONT_NAME] 13 | rcParams["font.weight"] = FONT_WEIGHT 14 | 15 | 16 | def plot_heatmap(matrix, labels=[], path=None): 17 | """Util function to plot confusion matrix""" 18 | cmap = matplotlib.colors.LinearSegmentedColormap.from_list("", ["#d75d5b", "#a7c3cd"]) 19 | fig, ax = plt.subplots(figsize=(15, 15)) 20 | ax.matshow(matrix, cmap=cmap, alpha=0.3) 21 | for i in range(matrix.shape[0]): 22 | for j in range(matrix.shape[1]): 23 | ax.text( 24 | x=j, 25 | y=i, 26 | s=f"{matrix[i, j]:.2f}", 27 | va="center", 28 | ha="center", 29 | # size="xx-large", 30 | ) 31 | 32 | plt.xticks(ticks=range(len(labels)), labels=labels) 33 | plt.yticks(ticks=range(len(labels)), labels=labels) 34 | plt.tight_layout() 35 | if path: 36 | plt.savefig(path) 37 | else: 38 | plt.show() 39 | plt.close() 40 | 41 | 42 | def plot_lines(x, y, xlim=None, ylim=None, labels=[], title=None, xlabel=None, ylabel=None, size=(), path=None): 43 | """Util function to plot lines""" 44 | plt.figure(figsize=size if size else (4, 1.5)) # width:20, height:3 45 | # plt.figure(figsize=size if size else (3, 2)) # width:20, height:3 46 | markers = [ 47 | "o", 48 | "v", 49 | "^", 50 | "<", 51 | ">", 52 | "1", 53 | "2", 54 | "3", 55 | "4", 56 | "8", 57 | "s", 58 | "p", 59 | "P", 60 | "*", 61 | "h", 62 | "H", 63 | "+", 64 | "x", 65 | "X", 66 | "D", 67 | "d", 68 | "|", 69 | "_", 70 | ] 71 | colors = [ 72 | "#a7c3cd", 73 | "#8a4444", 74 | "#524a47", 75 | "#d75d5b", 76 | "#c8c5c3", 77 | "#f5f0ed", 78 | "#edeef0", 79 | ] 80 | 81 | if np.shape(x)[0] == 1: 82 | x = np.ravel(x) 83 | 84 | if np.shape(x) and np.shape(x)[0] > 1 and not isinstance(x[0], str) and np.shape(x[0]) and np.shape(x[0])[0] >= 1: 85 | for idx, (x_values, y_values) in enumerate(zip(x, y)): 86 | plt.plot( 87 | x_values, 88 | y_values, 89 | color=colors[idx] if idx < len(colors) else None, 90 | # marker=markers[idx] if idx < len(markers) else None, 91 | label=labels[idx] if idx < len(labels) else "", 92 | ) 93 | else: 94 | for idx, values in enumerate(y): 95 | plt.plot( 96 | x, 97 | values, 98 | color=colors[idx] if idx < len(colors) else None, 99 | # marker=markers[idx] if idx < len(markers) else None, 100 | label=labels[idx] if idx < len(labels) else "", 101 | ) 102 | 103 | if len(labels) > 1: 104 | plt.legend(loc="lower right", ncol=2) 105 | 106 | if xlabel: 107 | plt.xlabel(xlabel, fontname=FONT_NAME, fontweight=FONT_WEIGHT) 108 | 109 | if ylabel: 110 | plt.ylabel(ylabel, fontname=FONT_NAME, fontweight=FONT_WEIGHT) 111 | 112 | if xlim: 113 | plt.xlim(xlim) 114 | 115 | if ylim: 116 | plt.ylim(ylim) 117 | 118 | if title: 119 | plt.title(title) 120 | 121 | _, end = plt.xlim() 122 | end = int(end) 123 | plt.xticks(np.arange(0, end + 1, max(1, int(end / 10)))) 124 | 125 | plt.tight_layout() 126 | if path: 127 | plt.savefig(path) 128 | else: 129 | plt.show() 130 | plt.close() 131 | 132 | 133 | def plot_bars(x, y, ylim=None, xlabel=None, ylabel=None, labels=[], title=None, path=None): 134 | """Util function to plot bars""" 135 | plt.figure(figsize=(30, 3)) # width:20, height:3 136 | width = 0.4 137 | fig, ax = plt.subplots() 138 | locs = np.arange(len(x)) # the label locations 139 | colors = [ 140 | "#d75d5b", 141 | "#a7c3cd", 142 | "#524a47", 143 | "#8a4444", 144 | "#c8c5c3", 145 | "#524a47", 146 | "#edeef0", 147 | ] 148 | 149 | for idx, values in enumerate(y): 150 | ax.bar( 151 | locs - (width / 2) if idx % 2 == 0 else locs + (width / 2), 152 | values, 153 | width, 154 | color=colors[idx] if idx <= len(colors) else None, 155 | label=labels[idx] if idx < len(labels) else "", 156 | ) 157 | 158 | ax.set_xticks(locs) 159 | ax.set_xticklabels(x, rotation=60) 160 | if labels: 161 | ax.legend() 162 | 163 | if xlabel: 164 | plt.xlabel(xlabel) 165 | 166 | if ylabel: 167 | plt.ylabel(ylabel) 168 | 169 | if ylim: 170 | ax.set_ylim(ylim) 171 | 172 | if title: 173 | plt.title(title) 174 | 175 | fig.tight_layout() 176 | if path: 177 | plt.savefig(path) 178 | else: 179 | plt.show() 180 | plt.close() 181 | 182 | 183 | def plot_lines_and_bars( 184 | x, 185 | lines, 186 | bars, 187 | ylim=None, 188 | xlabel=None, 189 | ylabel=None, 190 | second_x_axis=None, 191 | second_x_axis_label=None, 192 | labels=[], 193 | legend=[], 194 | colors_by_x=[], 195 | title=None, 196 | path=None, 197 | ): 198 | """Util function to plot lines""" 199 | plt.figure(figsize=(40, 3)) # width:20, height:3 200 | 201 | width = 0.4 202 | fig, ax = plt.subplots() 203 | locs = np.arange(len(x)) # the label locations 204 | colors = [ 205 | "#d75d5b", 206 | "#a7c3cd", 207 | "#f5f0ed", 208 | "#524a47", 209 | "#8a4444", 210 | "#edeef0", 211 | "#c8c5c3", 212 | ] 213 | 214 | for idx, values in enumerate(lines): 215 | ax.plot( 216 | x, 217 | values, 218 | color=colors[idx] if idx < len(colors) else None, 219 | label=labels[idx] if idx < len(labels) else "", 220 | ) 221 | 222 | for idx, values in enumerate(bars): 223 | if colors_by_x: 224 | ax.bar( 225 | locs, 226 | values, 227 | width if len(bars) > 1 else 1, 228 | color=colors_by_x, 229 | ) 230 | else: 231 | ax.bar( 232 | locs - (width / 2) if idx % 2 == 0 else locs + (width / 2), 233 | values, 234 | width if len(bars) > 1 else 1, 235 | color=colors[len(colors) - idx - 1] if len(colors) - idx - 1 >= 0 else None, 236 | label=labels[idx] if idx < len(labels) else "", 237 | ) 238 | 239 | patches = [] 240 | if legend: 241 | for label, color in legend.items(): 242 | patches.append(mpatches.Patch(color=color, label=label)) 243 | 244 | ax.set_xticks(locs) 245 | ax.set_xticklabels(x, rotation=60) 246 | 247 | if second_x_axis is not None: 248 | ax2 = ax.twiny() 249 | ax2.set_xlim(ax.get_xlim()) 250 | ax2.set_xticks(locs) 251 | ax2.set_xticklabels(second_x_axis, rotation=60) 252 | 253 | if second_x_axis_label: 254 | ax2.set_xlabel(second_x_axis_label) 255 | 256 | if patches: 257 | plt.legend(handles=patches) 258 | elif labels: 259 | plt.legend() 260 | 261 | if xlabel: 262 | ax.set_xlabel(xlabel) 263 | 264 | if ylabel: 265 | ax.set_ylabel(ylabel) 266 | 267 | if ylim: 268 | plt.ylim(ylim) 269 | 270 | if title: 271 | plt.title(title) 272 | 273 | plt.tight_layout() 274 | if path: 275 | plt.savefig(path) 276 | else: 277 | plt.show() 278 | plt.close() 279 | 280 | 281 | def plot_stacked_bars(x, y, y_placeholder=None, ylim=None, xlabel=None, ylabel=None, labels=[], title=None, path=None): 282 | plt.figure(figsize=(50, 10)) # width:20, height:3 283 | """Util function to plot stacker bars""" 284 | fig, ax = plt.subplots() 285 | width = 0.8 286 | colors = [ 287 | "#a7c3cd", 288 | "#8a4444", 289 | "#c8c5c3", 290 | "#f5f0ed", 291 | "#d75d5b", 292 | ] 293 | # hatches = ["/", "-", "+", ".", "*"] 294 | 295 | y_placeholder = np.sort(y_placeholder, axis=0)[::-1] if y_placeholder else None 296 | labels = [label for _, label in (sorted(zip(y, labels), key=lambda pair: np.sum(pair[0]))[::-1])] 297 | y = np.sort(y, axis=0)[::-1] 298 | 299 | if y_placeholder is not None: 300 | previous_stack = 0 301 | for i, stack in enumerate(y_placeholder): 302 | if i > 0: 303 | previous_stack += y_placeholder[i - 1] 304 | 305 | rects = ax.bar( 306 | x, 307 | stack, 308 | width, 309 | color="#edeef0", 310 | edgecolor="#524a47", 311 | linewidth=0.25, 312 | bottom=previous_stack, 313 | ) 314 | sum_y = [sum(val) for val in zip(*y)] 315 | # ax.bar_label(rects, labels=[f"{val:.2f}" for val in sum_y], padding=1) 316 | 317 | bottom_by_y = {} 318 | if y_placeholder is not None: 319 | for i, stack in enumerate(y_placeholder): 320 | if i == 0: 321 | bottom_by_y[i] = stack 322 | else: 323 | bottom_by_y[i] = stack + bottom_by_y[i - 1] 324 | else: 325 | for i, stack in enumerate(y): 326 | if i == 0: 327 | bottom_by_y[i] = stack 328 | else: 329 | bottom_by_y[i] = stack + bottom_by_y[i - 1] 330 | 331 | for i, values in enumerate(y): 332 | rects = ax.bar( 333 | x, 334 | values, 335 | width, 336 | bottom=bottom_by_y[i - 1] if i > 0 and bottom_by_y else 0, 337 | # hatch=hatches[i] if i < len(hatches) else None, 338 | color=colors[i] if i < len(colors) else None, 339 | label=labels[i] if labels else "", 340 | ) 341 | # ax.bar_label(rects, label_type="center", fmt="%.2f", padding=5) 342 | 343 | if labels: 344 | ax.legend() 345 | 346 | if xlabel: 347 | plt.xlabel(xlabel) 348 | 349 | if ylabel: 350 | plt.ylabel(ylabel) 351 | 352 | if ylim: 353 | ax.set_ylim(ylim) 354 | 355 | plt.xticks(rotation=60) 356 | if title: 357 | plt.title(title) 358 | 359 | fig.tight_layout() 360 | plt.tight_layout() 361 | 362 | if path: 363 | plt.savefig(path) 364 | else: 365 | plt.show() 366 | plt.close() 367 | 368 | 369 | def plot_stacked_bars_split( 370 | x, y_a, y_b, y_placeholder=None, ylim=None, xlabel=None, ylabel=None, labels=[], title=None, path=None 371 | ): 372 | """Util function to plot stacker bars""" 373 | plt.figure(figsize=(50, 3)) # width:50, height:3 374 | fig, ax = plt.subplots() 375 | width = 0.8 376 | colors = [ 377 | "#a7c3cd", 378 | "#8a4444", 379 | "#c8c5c3", 380 | "#f5f0ed", 381 | "#d75d5b", 382 | ] 383 | # hatches = ["/", "-", "+", ".", "*"] 384 | 385 | labels = [label for _, label in sorted(zip(y_a, labels), key=lambda pair: np.sum(pair[0]))[::-1]] 386 | y_a = np.sort(y_a, axis=0)[::-1] 387 | y_b = np.sort(y_b, axis=0)[::-1] 388 | x = np.sort(x, axis=0)[::-1] 389 | y_placeholder = np.sort(y_placeholder, axis=0)[::-1] if y_placeholder else None 390 | 391 | locs = np.arange(len(x)) # the label locations 392 | new_locs = np.array([2 * i for i in locs]) 393 | if y_placeholder is not None: 394 | previous_stack = 0 395 | for i, stack in enumerate(y_placeholder): 396 | if i > 0: 397 | previous_stack += y_placeholder[i - 1] 398 | 399 | rects1 = ax.bar( 400 | new_locs - (width / 2), 401 | stack, 402 | width, 403 | color="#edeef0", 404 | edgecolor="#524a47", 405 | linewidth=0.25, 406 | bottom=previous_stack, 407 | ) 408 | rects2 = ax.bar( 409 | new_locs + (width / 2), 410 | stack, 411 | width, 412 | color="#edeef0", 413 | edgecolor="#524a47", 414 | linewidth=0.25, 415 | bottom=previous_stack, 416 | ) 417 | sum_y_a = [sum(val) for val in zip(*y_a)] 418 | sum_y_b = [sum(val) for val in zip(*y_b)] 419 | # ax.bar_label(rects1, labels=[f"{val:.2f}" for val in sum_y_a], padding=1, rotation=60) 420 | # ax.bar_label(rects2, labels=[f"{val:.2f}" for val in sum_y_b], padding=1, rotation=60) 421 | 422 | bottom_by_y = {} 423 | bottom_by_y_a = {} 424 | bottom_by_y_b = {} 425 | if y_placeholder is not None: 426 | for i, stack in enumerate(y_placeholder): 427 | if i == 0: 428 | bottom_by_y[i] = stack 429 | else: 430 | bottom_by_y[i] = stack + bottom_by_y[i - 1] 431 | else: 432 | for i, (values_a, values_b) in enumerate(zip(y_a, y_b)): 433 | if i == 0: 434 | bottom_by_y_a[i] = np.array(values_a) 435 | bottom_by_y_b[i] = np.array(values_b) 436 | else: 437 | bottom_by_y_a[i] = np.array(values_a) + bottom_by_y_a[i - 1] 438 | bottom_by_y_b[i] = np.array(values_b) + bottom_by_y_b[i - 1] 439 | 440 | for i, (values_a, values_b) in enumerate(zip(y_a, y_b)): 441 | rects1 = ax.bar( 442 | new_locs - (width / 2), 443 | values_a, 444 | width, 445 | bottom=bottom_by_y[i - 1] 446 | if i > 0 and bottom_by_y 447 | else bottom_by_y_a[i - 1] 448 | if i > 0 and bottom_by_y_a 449 | else 0, 450 | # hatch=hatches[i] if i < len(hatches) else None, 451 | color=colors[i] if i < len(colors) else None, 452 | label=labels[i] if labels and i < len(labels) else None, 453 | ) 454 | rects2 = ax.bar( 455 | new_locs + (width / 2), 456 | values_b, 457 | width, 458 | bottom=bottom_by_y[i - 1] 459 | if i > 0 and bottom_by_y 460 | else bottom_by_y_b[i - 1] 461 | if i > 0 and bottom_by_y_b 462 | else 0, 463 | # hatch=hatches[i] if i < len(hatches) else None, 464 | color=colors[i] if i < len(colors) else None, 465 | # label=labels[i] if labels else "", 466 | ) 467 | # ax.bar_label(rects1, label_type="center", fmt="%.2f", padding=5) 468 | # ax.bar_label(rects2, label_type="center", fmt="%.2f", padding=5) 469 | 470 | ax.set_xticks(new_locs) 471 | ax.set_xticklabels(x, rotation=60) 472 | if labels: 473 | ax.legend() 474 | 475 | if xlabel: 476 | plt.xlabel(xlabel) 477 | 478 | if ylabel: 479 | plt.ylabel(ylabel) 480 | 481 | if ylim: 482 | ax.set_ylim(ylim) 483 | 484 | if title: 485 | plt.title(title) 486 | 487 | fig.tight_layout() 488 | plt.tight_layout() 489 | 490 | if path: 491 | plt.savefig(path) 492 | else: 493 | plt.show() 494 | plt.close() 495 | -------------------------------------------------------------------------------- /trustee/utils/rootpath.py: -------------------------------------------------------------------------------- 1 | """ 2 | This code was copied from the python-rootpath project linked below. 3 | We copied this snippet of code with no intention of stealing their code, 4 | but to fix an install issue due rootpath depending on codecov. 5 | 6 | To avoid further issues like this in the future, it seemed like a good idea 7 | to just incorporate the specific function we needed from the project. 8 | 9 | * python-rootpath: 10 | - https://github.com/grimen/python-rootpath 11 | * linked issues: 12 | - https://github.com/TrusteeML/trustee/issues/2 13 | - https://community.codecov.com/t/codecov-yanked-from-pypi-all-versions/4259/11 14 | """ 15 | # ========================================= 16 | # IMPORTS 17 | # -------------------------------------- 18 | 19 | import sys 20 | import os 21 | import re 22 | import six 23 | 24 | from os import path, listdir 25 | 26 | 27 | # ========================================= 28 | # CONSTANTS 29 | # -------------------------------------- 30 | 31 | DEFAULT_PATH = "." 32 | DEFAULT_ROOT_FILENAME_MATCH_PATTERN = ".git|requirements.txt" 33 | 34 | 35 | # ========================================= 36 | # FUNCTIONS 37 | # -------------------------------------- 38 | 39 | 40 | def detect(current_path=None, pattern=None): 41 | 42 | """ 43 | Find project root path from specified file/directory path, 44 | based on common project root file pattern. 45 | 46 | Examples: 47 | 48 | import rootpath 49 | 50 | rootpath.detect() 51 | rootpath.detect(__file__) 52 | rootpath.detect('./src') 53 | 54 | """ 55 | 56 | current_path = current_path or os.getcwd() 57 | current_path = path.abspath(path.normpath(path.expanduser(current_path))) 58 | pattern = pattern or DEFAULT_ROOT_FILENAME_MATCH_PATTERN 59 | 60 | if not path.isdir(current_path): 61 | current_path = path.dirname(current_path) 62 | 63 | def find_root_path(current_path, pattern=None): 64 | if isinstance(pattern, six.string_types): 65 | pattern = re.compile(pattern) 66 | 67 | detecting = True 68 | 69 | found_more_files = None 70 | found_root = None 71 | found_system_root = None 72 | 73 | file_names = None 74 | root_file_names = None 75 | 76 | while detecting: 77 | file_names = listdir(current_path) 78 | found_more_files = bool(len(file_names) > 0) 79 | 80 | if not found_more_files: 81 | detecting = False 82 | 83 | return None 84 | 85 | root_file_names = filter(pattern.match, file_names) 86 | root_file_names = list(root_file_names) 87 | 88 | found_root = bool(len(root_file_names) > 0) 89 | 90 | if found_root: 91 | detecting = False 92 | 93 | return current_path 94 | 95 | found_system_root = bool(current_path == path.sep) 96 | 97 | if found_system_root: 98 | return None 99 | 100 | system_root = sys.executable 101 | 102 | while os.path.split(system_root)[1]: 103 | system_root = os.path.split(system_root)[0] 104 | 105 | if current_path == system_root: 106 | return None 107 | 108 | current_path = path.abspath(path.join(current_path, "..")) 109 | 110 | return find_root_path(current_path, pattern) 111 | -------------------------------------------------------------------------------- /trustee/utils/tree.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | 3 | from copy import deepcopy 4 | 5 | from sklearn.tree._tree import TREE_LEAF, TREE_UNDEFINED, NODE_DTYPE 6 | 7 | 8 | def prune_index(dt, index, prune_level): 9 | """Prunes the given decision tree at the given index and returns the number of pruned nodes""" 10 | if index < 0: 11 | return 0 12 | 13 | left_idx = dt.tree_.children_left[index] 14 | right_idx = dt.tree_.children_right[index] 15 | 16 | # turn node into a leaf by "unlinking" its children 17 | dt.tree_.children_left[index] = TREE_LEAF if prune_level == 0 else TREE_UNDEFINED 18 | dt.tree_.children_right[index] = TREE_LEAF if prune_level == 0 else TREE_UNDEFINED 19 | 20 | # if there are children, visit them as well 21 | if left_idx != TREE_LEAF and right_idx != TREE_LEAF: 22 | prune_index(dt, left_idx, prune_level + 1) 23 | prune_index(dt, right_idx, prune_level + 1) 24 | 25 | 26 | def get_dt_dict(dt): 27 | """Iterates through the given Decision Tree to collect updated tree node structure""" 28 | children_left = dt.tree_.children_left 29 | children_right = dt.tree_.children_right 30 | dt_state = dt.tree_.__getstate__() 31 | nodes_arr = dt_state["nodes"] 32 | values_arr = dt_state["values"] 33 | 34 | idx_inc = 0 35 | nodes = [] 36 | 37 | def walk_tree(node, level, idx): 38 | """Recursively iterates through all nodes in given decision tree and returns them as a list.""" 39 | left = children_left[node] 40 | right = children_right[node] 41 | 42 | nonlocal idx_inc 43 | if left != right: # if not leaf node 44 | idx_inc += 1 45 | left = walk_tree(left, level + 1, idx_inc) 46 | idx_inc += 1 47 | right = walk_tree(right, level + 1, idx_inc) 48 | 49 | nodes.append( 50 | { 51 | "idx": idx, 52 | "node": node, 53 | "left": left, 54 | "right": right, 55 | "level": level, 56 | "node_tuple": nodes_arr[node], 57 | "values_tuple": values_arr[node], 58 | } 59 | ) 60 | 61 | return idx 62 | 63 | walk_tree(0, 0, idx_inc) 64 | 65 | max_depth = 0 66 | node_values = [] 67 | node_ndarray = np.array([], dtype=NODE_DTYPE) 68 | 69 | for node in sorted(nodes, key=lambda x: x["idx"]): 70 | if node["level"] > max_depth: 71 | max_depth = node["level"] 72 | 73 | node_tuple = (node["left"], node["right"], *tuple(node["node_tuple"])[2:]) 74 | node_ndarray = np.append(node_ndarray, np.array([node_tuple], dtype=NODE_DTYPE)) 75 | node_values.append(node["values_tuple"]) 76 | 77 | value_ndarray = np.array(node_values, dtype=np.float64) 78 | 79 | dt_dict = { 80 | "max_depth": max_depth, 81 | "node_count": len(node_ndarray), 82 | "nodes": node_ndarray, 83 | "values": value_ndarray, 84 | } 85 | 86 | return dt_dict 87 | 88 | 89 | def get_dt_info(dt): 90 | """Iterates through the given Decision Tree to collect relevant information.""" 91 | children_left = dt.tree_.children_left 92 | children_right = dt.tree_.children_right 93 | features = dt.tree_.feature 94 | thresholds = dt.tree_.threshold 95 | values = dt.tree_.value 96 | samples = dt.tree_.n_node_samples 97 | impurity = dt.tree_.impurity 98 | 99 | splits = [] 100 | features_used = {} 101 | 102 | def walk_tree(node, level, path): 103 | """Recursively iterates through all nodes in given decision tree and returns them as a list.""" 104 | if children_left[node] == children_right[node]: # if leaf node 105 | node_class = np.argmax(values[node][0]) if len(np.array(values[node][0])) > 1 else values[node][0][0] 106 | node_prob = ( 107 | (values[node][0][node_class] / np.sum(values[node][0])) * 100 108 | if np.array(values[node][0]).ndim > 1 109 | else 0 110 | ) 111 | return [ 112 | { 113 | "level": level, 114 | "path": path, 115 | "class": node_class, 116 | "prob": node_prob, 117 | "samples": samples[node], 118 | } 119 | ] 120 | 121 | feature = features[node] 122 | threshold = thresholds[node] 123 | left = children_left[node] 124 | right = children_right[node] 125 | 126 | if feature not in features_used: 127 | features_used[feature] = {"count": 0, "samples": 0} 128 | 129 | features_used[feature]["count"] += 1 130 | features_used[feature]["samples"] += samples[node] 131 | 132 | splits.append( 133 | { 134 | "idx": node, 135 | "level": level, 136 | "feature": feature, 137 | "threshold": threshold, 138 | "samples": samples[node], 139 | "values": values[node], 140 | "gini_split": (impurity[left], impurity[right]), 141 | "data_split": (np.sum(values[left]), np.sum(values[right])), 142 | "data_split_by_class": [ 143 | (c_left, c_right) for (c_left, c_right) in zip(values[left][0], values[right][0]) 144 | ], 145 | } 146 | ) 147 | 148 | return walk_tree(left, level + 1, path + [(node, feature, "<=", threshold)]) + walk_tree( 149 | right, level + 1, path + [(node, feature, ">", threshold)] 150 | ) 151 | 152 | branches = walk_tree(0, 0, []) 153 | return features_used, splits, branches 154 | 155 | 156 | def top_k_prune(dt, top_k, max_impurity=0.1): 157 | """Prunes a given decision tree down to its top-k branches, sorted by number of samples covered""" 158 | _, nodes, branches = get_dt_info(dt) 159 | top_branches = sorted(branches, key=lambda p: p["samples"], reverse=True)[:top_k] 160 | prunned_dt = deepcopy(dt) 161 | 162 | nodes_to_keep = set({}) 163 | for branch in top_branches: 164 | for node, _, _, _ in branch["path"]: 165 | if dt.tree_.impurity[node] > max_impurity: 166 | nodes_to_keep.add(node) 167 | 168 | for node in nodes: 169 | if node["idx"] not in nodes_to_keep: 170 | prune_index(prunned_dt, node["idx"], 0) 171 | 172 | # update classifier with prunned model 173 | prunned_dt.tree_.__setstate__(get_dt_dict(prunned_dt)) 174 | 175 | return prunned_dt 176 | --------------------------------------------------------------------------------