├── LICENSE ├── README.md ├── src ├── README.md ├── bagging_variation.m ├── classifier_test.m ├── classifier_train.m ├── follow_the_leader.m ├── learn.m ├── learn_nse.m ├── sea.m ├── smote.m ├── smv.m ├── stats.m ├── swmv.m └── test_then_train.m └── tests ├── README.md ├── test_ftl.m ├── test_learn.m ├── test_learn_nse.m ├── test_smv.m ├── test_swmv.m └── test_ttt.m /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 | IncrementalLearning 2 | =================== 3 | 4 | I am planning on providing some implementations of Learn++ and other incremental learning algorithms in Matlab. Please direct any questions, comments, or suggestions to . More details to come. 5 | -------------------------------------------------------------------------------- /src/README.md: -------------------------------------------------------------------------------- 1 | IncrementalLearning 2 | =================== 3 | 4 | I am planning on providing some implementations of the Learn++ algorithm in Matlab. Please direct any questions, comments, or suggestions to . More details to come. 5 | 6 | 7 | 8 | About the functions 9 | =================== 10 | I have added documentation to all the files. Please use `help(myFun)` (where `myFun` is the function you are interested in) in the Matlab commandline for a detailed description on how to use the function. Also refer to the unit tests for examples. 11 | 12 | 13 | References 14 | =================== 15 | 1. R. Polikar, L. Udpa, S. Udpa, and V. Honavar, "Learn++: An incremental learning algorithm for supervised neural networks," IEEE Transactions on System, Man and Cybernetics (C), Special Issue on Knowledge Management, vol. 31, no. 4, pp. 497-508, 2001. 16 | 2. R. Elwell and R. Polikar, "Incremental Learning of Concept Drift in Nonstationary Environments" IEEE Transactions on Neural Networks, vol. 22, no. 10, pp. 1517-1531 17 | 3. G. Ditzler and R. Polikar, "Incremental learning of concept drift from streaming imbalanced data," in IEEE Transactions on Knowledge & Data Engineering, 2013, vol. 25, no. 10, pp. 2283-2301. 18 | 4. N. V. Chawla, K. W. Bowyer, T. E. Moore and P. Kegelmeyer, "SMOTE: Synthetic Minority Over-Sampling Technique," Journal of Artificial Intelligence Research, 16, 321-357, 2002. 19 | -------------------------------------------------------------------------------- /src/bagging_variation.m: -------------------------------------------------------------------------------- 1 | function sub_ensemble = bagging_variation(data, labels, n_classifiers, minority_class, base_classifier) 2 | % sub_ensemble = bagging_variation(data, labels, ... 3 | % n_classifiers, minority_class, base_classifier) 4 | % 5 | % @data - data matrix in n_observations by number of features 6 | % matrix 7 | % @labels - labels in an n_observations by one vector 8 | % @n_classifiers - number of classifiers to generate in the 9 | % subensemble 10 | % @minority_class - integer specifying which class in the 11 | % prediction problem is the minority 12 | % @base_classifier - base classification algorithm used in 13 | % CLASSIFIER_TRAIN.m 14 | % @sub_ensemble - cell array containing the sub ensemble of 15 | % classifiers. 16 | % 17 | % This function implements the bagging variation algorithm used in 18 | % Learn++.NIE. 19 | % 20 | % Cite: 21 | % 1) G. Ditzler and R. Polikar, "Incremental learning of concept drift 22 | % from streaming imbalanced data," in IEEE Transactions on Knowledge 23 | % & Data Engineering, 2012, accepted. 24 | % 25 | % @Author: Gregory Ditzler (gregory.ditzler@gmail.com) 26 | 27 | 28 | % bagging_variation.m 29 | % Copyright (C) 2013 Gregory Ditzler 30 | % 31 | % This program is free software: you can redistribute it and/or modify 32 | % it under the terms of the GNU General Public License as published by 33 | % the Free Software Foundation, either version 3 of the License, or 34 | % (at your option) any later version. 35 | % 36 | % This program is distributed in the hope that it will be useful, 37 | % but WITHOUT ANY WARRANTY; without even the implied warranty of 38 | % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 39 | % GNU General Public License for more details. 40 | % 41 | % You should have received a copy of the GNU General Public License 42 | % along with this program. If not, see . 43 | 44 | 45 | negative_indices = find(labels ~= minority_class); 46 | positive_indices = find(labels == minority_class); 47 | sub_ensemble = cell(n_classifiers, 1); 48 | 49 | parfor k = 1:n_classifiers 50 | index = negative_indices(randi(numel(negative_indices), 1, ... 51 | floor(numel(labels)/n_classifiers))); 52 | data_k = [data(index, :); data(positive_indices, :)]; 53 | label_k = [labels(index); labels(positive_indices)]; 54 | sub_ensemble{k} = classifier_train(base_classifier, data_k, label_k); 55 | end -------------------------------------------------------------------------------- /src/classifier_test.m: -------------------------------------------------------------------------------- 1 | function predictions = classifier_test(model, data) 2 | % predictions = CLASSIFIER_TEST(model, data) 3 | % @model - you get this structure from CLASSIFIER_TRAIN.m 4 | % @data - matrix of data. n_observations by n_features 5 | % @predictions - vector of predictions made by the classifier 6 | % trained in the CLASSIFIER_TRAIN.m function 7 | % 8 | % Test a classifier on data. 9 | % 10 | % @Author: Gregory Ditzler (gregory.ditzler@gmail.com) 11 | % 12 | % See also 13 | % classifier_train.m 14 | 15 | % classifier_test.m 16 | % Copyright (C) 2013 Gregory Ditzler 17 | % 18 | % This program is free software: you can redistribute it and/or modify 19 | % it under the terms of the GNU General Public License as published by 20 | % the Free Software Foundation, either version 3 of the License, or 21 | % (at your option) any later version. 22 | % 23 | % This program is distributed in the hope that it will be useful, 24 | % but WITHOUT ANY WARRANTY; without even the implied warranty of 25 | % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 26 | % GNU General Public License for more details. 27 | % 28 | % You should have received a copy of the GNU General Public License 29 | % along with this program. If not, see . 30 | 31 | % You can easily modify this file to include new baseclassifiers. This file 32 | % must have lines added for creating the classifier that you are interested 33 | % in. You also need to modify CLASSIFIER_TRAIN.M to implement the evaluation 34 | % of the classifier on a test set. 35 | 36 | 37 | switch model.type 38 | case 'CART' 39 | predictions_raw = model.classifier(data); 40 | % cart returns a cell arrary of strings; convert the predictions to a 41 | % vector of indices for the class being predicted. 42 | predictions = zeros(length(predictions_raw),1); 43 | for m = 1:numel(predictions) 44 | predictions(m) = str2double(predictions_raw{m}); 45 | end 46 | end -------------------------------------------------------------------------------- /src/classifier_train.m: -------------------------------------------------------------------------------- 1 | function model = classifier_train(model, data, labels) 2 | % model = CLASSIFIER_TRAIN(model, data) 3 | % @model - structure. you must set the type field. Only CART is 4 | % implemented at the moment. 5 | % (manditory fields) 6 | % > .type - 'CART' 7 | % (optional fields) 8 | % > .prune - see CLASSREGTREE.m 9 | % > .minleaf - see CLASSREGTREE.m 10 | % > .mergeleaves - see CLASSREGTREE.m 11 | % > .surrogate - see CLASSREGTREE.m 12 | % @data - matrix of data. n_observations by n_features 13 | % @labels - vector of labels made by the classifier 14 | % @model - update model structure with a trained classifier 15 | % 16 | % Train a classifier on labeled data. 17 | % 18 | % @Author: Gregory Ditzler (gregory.ditzler@gmail.com) 19 | % 20 | % See also 21 | % classifier_test.m 22 | 23 | % classifier_train.m 24 | % Copyright (C) 2013 Gregory Ditzler 25 | % 26 | % This program is free software: you can redistribute it and/or modify 27 | % it under the terms of the GNU General Public License as published by 28 | % the Free Software Foundation, either version 3 of the License, or 29 | % (at your option) any later version. 30 | % 31 | % This program is distributed in the hope that it will be useful, 32 | % but WITHOUT ANY WARRANTY; without even the implied warranty of 33 | % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 34 | % GNU General Public License for more details. 35 | % 36 | % You should have received a copy of the GNU General Public License 37 | % along with this program. If not, see . 38 | 39 | 40 | % You can easily modify this file to include new baseclassifiers. This file 41 | % must have lines added for creating the classifier that you are interested 42 | % in. You also need to modify CLASSIFIER_TEST.M to implement the evaluation 43 | % of the classifier on a test set. 44 | 45 | [n_observations, n_features] = size(data); 46 | if length(labels) ~= n_observations 47 | error(['CLASSIFIER_TRAIN.M :: Data obervations not equal to the number',... 48 | ' of labels']) 49 | end 50 | 51 | switch model.type 52 | case 'CART' 53 | % the CART classifier is the base model 54 | model.n_features = n_features; 55 | model.method = 'classification'; 56 | 57 | if isfield(model, 'prune') == 0 58 | % compute the full tree and the optimal sequence of pruned 59 | % subtrees, or 'off' for the full tree without pruning. 60 | model.prune = 'on'; 61 | end 62 | if isfield(model, 'minleaf') == 0 63 | % minimal number of observations per tree leaf 64 | model.minleaf = 1; 65 | end 66 | if isfield(model, 'mergeleaves') == 0 67 | % rge leaves that originate from the same parent node and give the 68 | % sum of risk values greater or equal to the risk associated with the 69 | % parent node. 70 | model.mergeleaves = 'on'; 71 | end 72 | if isfield(model, 'surrogate') == 0 73 | % recall that surrogate tree are for missing data. the classregtree 74 | % impementation is generally much slower when this is set to on. and 75 | % it is on bydefault. lets turn it off! 76 | model.surrogate = 'off'; 77 | end 78 | 79 | model.classifier = classregtree(data, labels, ... 80 | 'method', model.method,... 81 | 'prune', model.prune,... 82 | 'minleaf', model.minleaf,... 83 | 'mergeleaves', model.mergeleaves,... 84 | 'surrogate', model.surrogate); 85 | 86 | end -------------------------------------------------------------------------------- /src/follow_the_leader.m: -------------------------------------------------------------------------------- 1 | function [net,f_measure,g_mean,precision,recall,err] = follow_the_leader(net, data_train, labels_train, data_test, labels_test) 2 | % [net,f_measure,g_mean,precision,recall,err] = follow_the_leader(net, ... 3 | % data_train, labels_train, ... 4 | % data_test, labels_test, ... 5 | % smote_params) 6 | % 7 | % @net - initialized structure. you must initialize 8 | % net.mclass - number of classes 9 | % net.base_classifier - you should set this to be model.type 10 | % which is submitted to CLASSIFIER_TRAIN.m 11 | % net.n_classifiers - number of classifiers to keep in the pool 12 | % @data_train - cell array of training data. each entry should 13 | % have a n_oberservation by n_feature matrix 14 | % @labels_train - cell array of class labels 15 | % @data_test - cell array of training data. each entry should 16 | % have a n_oberservation by n_feature matrix 17 | % @labels_test - cell array of class labels 18 | % 19 | % Implementation of model selection scheme that uses only the 20 | % best perfoming classifier. 21 | % 22 | % @Author: Gregory Ditzler (gregory.ditzler@gmail.com) 23 | % 24 | % See also 25 | % CLASSIFIER_TRAIN.m CLASSIFIER_TEST.m 26 | 27 | % follow_the_leader.m 28 | % Copyright (C) 2013 Gregory Ditzler 29 | % 30 | % This program is free software: you can redistribute it and/or modify 31 | % it under the terms of the GNU General Public License as published by 32 | % the Free Software Foundation, either version 3 of the License, or 33 | % (at your option) any later version. 34 | % 35 | % This program is distributed in the hope that it will be useful, 36 | % but WITHOUT ANY WARRANTY; without even the implied warranty of 37 | % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 38 | % GNU General Public License for more details. 39 | % 40 | % You should have received a copy of the GNU General Public License 41 | % along with this program. If not, see . 42 | 43 | 44 | 45 | 46 | n_timestamps = length(data_train); % total number of time stamps 47 | net.classifiers = cell(net.n_classifiers,1); % classifiers 48 | net.type = 'follow-the-leader'; 49 | 50 | f_measure = zeros(n_timestamps, net.mclass); 51 | g_mean = zeros(n_timestamps, 1); 52 | recall = zeros(n_timestamps, net.mclass); 53 | precision = zeros(n_timestamps, net.mclass); 54 | err = zeros(n_timestamps, 1); 55 | 56 | if net.n_classifiers == -1 57 | net.n_classifiers = n_timestamps; 58 | end 59 | 60 | for ell = 1:n_timestamps 61 | % get the training data for the 't'th round 62 | data_train_t = data_train{ell}; 63 | labels_train_t = labels_train{ell}; 64 | data_test_t = data_test{ell}; 65 | labels_test_t = labels_test{ell}; 66 | index = mod(ell-1, net.n_classifiers) + 1; 67 | 68 | net.classifiers{index} = classifier_train(... 69 | net.base_classifier, ... 70 | data_train_t, ... 71 | labels_train_t); 72 | if ell < net.n_classifiers 73 | T = ell; 74 | else 75 | T = net.n_classifiers; 76 | end 77 | y = decision_ensemble(net, data_train_t, labels_train_t, T); 78 | e = zeros(T,1); 79 | for t = 1:T 80 | [~,~,~,~, e(t)] = stats(labels_train_t, y(:,t), net.mclass); 81 | end 82 | [~, i] = min(e); 83 | h = classifier_test(net.classifiers{i}, data_test_t); 84 | [f_measure(ell,:),g_mean(ell),recall(ell,:),precision(ell,:),... 85 | err(ell)] = stats(labels_test_t, h, net.mclass); 86 | end 87 | 88 | 89 | 90 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 91 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 92 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 93 | %%% AUXILARY FUNCTIONS 94 | function y = decision_ensemble(net, data, labels, n_experts) 95 | y = zeros(numel(labels), n_experts); 96 | for k = 1:n_experts 97 | y(:, k) = classifier_test(net.classifiers{k}, data); 98 | end 99 | -------------------------------------------------------------------------------- /src/learn.m: -------------------------------------------------------------------------------- 1 | function [net,errs] = learn(net, data_train, labels_train, data_test, labels_test) 2 | % [net,errs] = learn(net, data_train, labels_train, ... 3 | % data_test, labels_test) 4 | % 5 | % @net - initialized structure. you must initialize 6 | % net.iterations 7 | % net.base_classifier - you should set this to be model.type 8 | % which is submitted to CLASSIFIER_TRAIN.m 9 | % net.mclass - number of classes 10 | % @data_train - training data in a cell array. each entry should 11 | % have a n_oberservation by n_feature matrix 12 | % @labels_train - cell array of class labels 13 | % @data_test - test data in a matrix. the size of the matrix should 14 | % be n_oberservation by n_feature matrix 15 | % @labels_test -labels to the test data 16 | % @errs - error of the Learn++ on the testing data set. error is 17 | % measured at each addition of a new classifier. 18 | % 19 | % Implementation of Learn++. 20 | % 21 | % Cite: 22 | % 1) R. Polikar, L. Udpa, S. Udpa, and V. Honavar, "Learn++: An 23 | % incremental learning algorithm for supervised neural networks," 24 | % IEEE Transactions on System, Man and Cybernetics (C), Special 25 | % Issue on Knowledge Management, vol. 31, no. 4, pp. 497-508, 2001. 26 | % 27 | % See also 28 | % CLASSIFIER_TRAIN.m CLASSIFIER_TEST.m 29 | 30 | % learn.m 31 | % Copyright (C) 2013 Gregory Ditzler 32 | % 33 | % This program is free software: you can redistribute it and/or modify 34 | % it under the terms of the GNU General Public License as published by 35 | % the Free Software Foundation, either version 3 of the License, or 36 | % (at your option) any later version. 37 | % 38 | % This program is distributed in the hope that it will be useful, 39 | % but WITHOUT ANY WARRANTY; without even the implied warranty of 40 | % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 41 | % GNU General Public License for more details. 42 | % 43 | % You should have received a copy of the GNU General Public License 44 | % along with this program. If not, see . 45 | 46 | 47 | Tk = net.iterations; % number of classifiers to generate 48 | K = length(data_train); % number of data sets 49 | net.classifiers = cell(Tk*K, 1); % cell array with total number of classifiers 50 | net.beta = zeros(Tk*K, 1); % beta will set the classifier weights 51 | c_count = 0; % keep track of the number of classifiers at each time 52 | errs = zeros(Tk*K, 1); % prediction errors on the test data set 53 | 54 | % run learn++ on the data 55 | for k = 1:K 56 | 57 | % obtain the latest data set and initialize the weights over the 58 | % instances to form a uniform distribution 59 | data_train_k = data_train{k}; 60 | labels_train_k = labels_train{k}; 61 | D = ones(numel(labels_train_k), 1)/numel(labels_train_k); 62 | 63 | % original paper says to modify D if prior knowledge is available. we can 64 | % modify the distribution weights if we already have a classifier 65 | % ensemble. 66 | if k > 1 67 | predictions = classify_ensemble(net, data_train_k, labels_train_k, ... 68 | c_count); % predict on the training data 69 | epsilon_kt = sum(D(predictions ~= labels_train_k)); % error on D 70 | beta_kt = epsilon_kt/(1-epsilon_kt); % normalized error on D 71 | D(predictions == labels_train_k) = beta_kt * D(predictions == labels_train_k); 72 | end 73 | 74 | for t = 1:Tk 75 | % update the classifier count 76 | c_count = c_count + 1; 77 | 78 | % step 1 - make sure we are working with a probability distribution. 79 | D = D / sum(D); 80 | 81 | % step 2 - grab a random sample of data indices with replacement from 82 | % the probability distribution D 83 | index = randsample(1:numel(D), numel(D), true, D); 84 | 85 | % step 3 - generate a new classifier on the data sampled from D. 86 | net.classifiers{c_count} = classifier_train(... 87 | net.base_classifier, ... 88 | data_train_k(index, :), ... 89 | labels_train_k(index)); 90 | 91 | % step 4 - test the latest classifier on ALL of the data not just the 92 | % data sampled from D, and compute the error according to the 93 | % probability distribution. then compute beta 94 | y = classifier_test(net.classifiers{c_count}, data_train_k); 95 | epsilon_kt = sum(D(y ~= labels_train_k)); 96 | net.beta(c_count) = epsilon_kt/(1-epsilon_kt); 97 | 98 | % step 5 - get the ensemble decision computed with c_count classifiers 99 | % in the ensemble. compute the error on the probability distribution on 100 | % the composite hypothesis. 101 | predictions = classify_ensemble(net, data_train_k, labels_train_k, ... 102 | c_count); 103 | E_kt = sum(D(predictions ~= labels_train_k)); 104 | if E_kt > 0.5 105 | % rather than remove remove existing classifier; null the result out 106 | % by forcing the loss to be equal to 1/2 which is the worst possible 107 | % loss. feel free to modify the code to go back an iteration. 108 | E_kt = 0.5; 109 | end 110 | 111 | % step 6 - compute the normalized error of the compsite hypothesis and 112 | % update the weights over the training instances in the kth batch. 113 | Bkt = E_kt / (1 - E_kt); 114 | D(predictions == labels_train_k) = Bkt * D(predictions == labels_train_k); 115 | D = D / sum(D); 116 | 117 | % make some predictions on the testing data set. 118 | [predictions,posterior] = classify_ensemble(net, data_test, ... 119 | labels_test, c_count); 120 | errs(c_count) = sum(predictions ~= labels_test)/numel(labels_test); 121 | end 122 | 123 | 124 | end 125 | 126 | 127 | 128 | 129 | 130 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 131 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 132 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 133 | %%% AUXILARY FUNCTIONS 134 | function [predictions,posterior] = classify_ensemble(net, data, labels, lims) 135 | n_experts = lims; 136 | weights = log(1./net.beta(1:lims)); 137 | p = zeros(numel(labels), net.mclass); 138 | for k = 1:n_experts 139 | y = classifier_test(net.classifiers{k}, data); 140 | 141 | % this is inefficient, but it does the job 142 | for m = 1:numel(y) 143 | p(m,y(m)) = p(m,y(m)) + weights(k); 144 | end 145 | end 146 | [~,predictions] = max(p'); 147 | predictions = predictions'; 148 | posterior = p./repmat(sum(p,2),1,net.mclass); 149 | -------------------------------------------------------------------------------- /src/learn_nse.m: -------------------------------------------------------------------------------- 1 | function [net,f_measure,g_mean,precision,recall,err] = learn_nse(net, data_train, labels_train, data_test, ... 2 | labels_test, smote_params) 3 | % [net,f_measure,g_mean,precision,recall,err] = learn_nse(net, ... 4 | % data_train, labels_train, ... 5 | % data_test, labels_test, ... 6 | % smote_params) 7 | % 8 | % @net - initialized structure. you must initialize 9 | % net.a - sigmoid slope (try 0.5) 10 | % net.b - sigmoid cutoff (try 10) 11 | % net.threshold - small error threshold (try 0.01) 12 | % net.mclass - number of classes 13 | % net.base_classifier - you should set this to be model.type 14 | % which is submitted to CLASSIFIER_TRAIN.m 15 | % @data_train - cell array of training data. each entry should 16 | % have a n_oberservation by n_feature matrix 17 | % @labels_train - cell array of class labels 18 | % @data_test - cell array of training data. each entry should 19 | % have a n_oberservation by n_feature matrix 20 | % @labels_test - cell array of class labels 21 | % @smote_params - optional structure for implementing learn++.cds 22 | % smote_params.minority_class - minority class (scalar) 23 | % smote_params.k - see SMOTE.m 24 | % smote_params.N - see SMOTE.m 25 | % 26 | % Implementation of Learn++.NSE and Learn++.CDS. If @smote_params 27 | % is specified then the implementation is Learn++.CDS 28 | % 29 | % Cite: 30 | % 1) Elwell R. and Polikar R., "Incremental Learning of Concept Drift 31 | % in Nonstationary Environments" IEEE Transactions on Neural Networks, 32 | % vol. 22, no. 10, pp. 1517-1531 33 | % 2) G. Ditzler and R. Polikar, "Incremental learning of concept drift 34 | % from streaming imbalanced data," in IEEE Transactions on Knowledge 35 | % & Data Engineering, 2012, accepted. 36 | % 37 | % @Author: Gregory Ditzler (gregory.ditzler@gmail.com) 38 | % 39 | % See also 40 | % SMOTE.m CLASSIFIER_TRAIN.m CLASSIFIER_TEST.m 41 | 42 | 43 | 44 | % learn_nse.m 45 | % Copyright (C) 2013 Gregory Ditzler 46 | % 47 | % This program is free software: you can redistribute it and/or modify 48 | % it under the terms of the GNU General Public License as published by 49 | % the Free Software Foundation, either version 3 of the License, or 50 | % (at your option) any later version. 51 | % 52 | % This program is distributed in the hope that it will be useful, 53 | % but WITHOUT ANY WARRANTY; without even the implied warranty of 54 | % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 55 | % GNU General Public License for more details. 56 | % 57 | % You should have received a copy of the GNU General Public License 58 | % along with this program. If not, see . 59 | 60 | 61 | if nargin < 5 62 | error('LEARN_NSE :: You need to specify all the required inputs. ') 63 | end 64 | if nargin == 5 65 | smote_params = {}; 66 | smote_on =false; 67 | net.type = 'learn++.nse'; 68 | end 69 | if nargin == 6 70 | smote_on = true; 71 | net.type = 'learn++.cds'; 72 | end 73 | 74 | n_timestamps = length(data_train); % total number of time stamps 75 | net.classifiers = {}; % classifiers 76 | net.w = []; % weights 77 | net.initialized = false;% set to false 78 | net.t = 1; % track the time of learning 79 | net.classifierweigths = {}; % array of classifier weights 80 | 81 | f_measure = zeros(n_timestamps, net.mclass); 82 | g_mean = zeros(n_timestamps, 1); 83 | recall = zeros(n_timestamps, net.mclass); 84 | precision = zeros(n_timestamps, net.mclass); 85 | err = zeros(n_timestamps, 1); 86 | 87 | 88 | for ell = 1:n_timestamps 89 | 90 | % get the training data for the 't'th round 91 | data_train_t = data_train{ell}; 92 | labels_train_t = labels_train{ell}; 93 | data_test_t = data_test{ell}; 94 | labels_test_t = labels_test{ell}; 95 | 96 | if smote_on == true 97 | % add learn++.cds functionality here 98 | syn_data = smote(... 99 | data_train_t(labels_train_t == smote_params.minority_class,:), ... 100 | smote_params.k, ... 101 | smote_params.N); 102 | data_train_t = [data_train_t; syn_data]; 103 | labels_train_t = [labels_train_t;... 104 | ones(size(syn_data,1),1) * smote_params.minority_class]; 105 | i = randperm(numel(labels_train_t)); 106 | labels_train_t = labels_train_t(i); 107 | data_train_t = data_train_t(i, :); 108 | end 109 | 110 | % has the 111 | if net.initialized == false, 112 | net.beta = []; 113 | end 114 | 115 | mt = size(data_train_t,1); % numnber of training examples 116 | Dt = ones(mt,1)/mt; % initialize instance weight distribution 117 | 118 | if net.initialized==1, 119 | % STEP 1: Compute error of the existing ensemble on new data 120 | predictions = classify_ensemble(net, data_train_t, labels_train_t); 121 | Et = sum((predictions~=labels_train_t)/mt); 122 | Bt = Et/(1-Et); % this is suggested in Metin's IEEE Paper 123 | if Bt==0, Bt = 1/mt; end; % clip 124 | 125 | % update and normalize the instance weights 126 | Dt(predictions==labels_train_t) = Dt(predictions==labels_train_t) * Bt; 127 | Dt = Dt/sum(Dt); 128 | end 129 | 130 | % STEP 3: New classifier 131 | net.classifiers{end + 1} = classifier_train(... 132 | net.base_classifier, ... 133 | data_train_t, ... 134 | labels_train_t); 135 | 136 | % STEP 4: Evaluate all existing classifiers on new data 137 | t = size(net.classifiers,2); 138 | y = decision_ensemble(net, data_train_t, labels_train_t, t); 139 | 140 | for k = 1:net.t 141 | epsilon_tk = sum(Dt(y(:, k) ~= labels_train_t)); 142 | 143 | if (k0.5) 144 | epsilon_tk = 0.5; 145 | elseif (k==net.t)&&(epsilon_tk>0.5) 146 | % try generate a new classifier 147 | net.classifiers{k} = classifier_train(... 148 | net.base_classifier, ... 149 | data_train_t, ... 150 | labels_train_t); 151 | epsilon_tk = sum(Dt(y(:, k) ~= labels_train_t)); 152 | epsilon_tk(epsilon_tk > 0.5) = 0.5; % we tried; clip the loss 153 | end 154 | net.beta(net.t,k) = epsilon_tk / (1-epsilon_tk); 155 | end 156 | 157 | % compute the classifier weights 158 | if net.t==1, 159 | if net.beta(net.t,net.t). 26 | 27 | 28 | T = size(data, 1); 29 | SMOTEd = []; 30 | 31 | % If N is less than 100%, randomize the minority class samples as only a 32 | % random percent of them will be SMOTEd 33 | if N < 100, 34 | T = round((N/100)*T); 35 | N = 100; 36 | end 37 | 38 | for i = 1:T, 39 | nnarray = []; 40 | synthetic = []; 41 | 42 | % determine the euclidean distance between the current minority sample 43 | % and reset of the other minority samples. then sort them in ascending 44 | % order 45 | for j = 1:T, 46 | if i ~= j 47 | euclid_dist(j,:) = data(i,:) - data(j,:); 48 | else 49 | % ignore the sample we are currently at from further calculations. 50 | euclid_dist(j,:) = inf * ones(1, size(data,2)); 51 | end 52 | end 53 | euclid_dist = sqrt(sum(euclid_dist.^2,2)); 54 | euclid_dist2 = sort(euclid_dist,'ascend'); 55 | 56 | % if we a really dealing with an imbalanced data set we may not have 57 | % enough samples to reduce to k nearest neighbors; instead grab all of 58 | % them 59 | if length(euclid_dist2)<=k, 60 | knn = euclid_dist2; 61 | k = length(euclid_dist2); 62 | else 63 | knn = euclid_dist2(1:k); 64 | end 65 | 66 | % determine the k-nearest neighbors to the minority sample that we are 67 | % interested in. 68 | for j = 1:length(euclid_dist), 69 | if sum(euclid_dist(j)==knn) 70 | % the current distance in euclid_dist is a nearest neigbor of 71 | % the minority sample. so nnarray will have the indices of the 72 | % nearest neighbors in the minority instance array. 73 | nnarray(end+1) = j; 74 | end 75 | end 76 | 77 | % generate the synthetic samples 78 | newindex = 1; % keeps a count of number of synthetic samples generated 79 | N1 = round(N/100); % DO NOT OVERWRITE N!!!! 80 | 81 | while N1~=0, 82 | % Choose a random number between 1 and k, call it nn. This step 83 | % chooses one of the k nearest neighbors of i 84 | nn = round((k-1)*rand+1); % perform a linear conversion to scale the 85 | % nn paramter between 1 and k 86 | gap = rand; 87 | dif = data(nnarray(nn), :) - data(i, :); 88 | synthetic(newindex,:) = data(i, :) + gap*dif; 89 | 90 | newindex = newindex+1; 91 | N1 = N1-1; 92 | end 93 | SMOTEd = [SMOTEd; synthetic]; 94 | clear euclid_dist euclid_dist2 N1 nnarray synthetic nnarray 95 | end -------------------------------------------------------------------------------- /src/smv.m: -------------------------------------------------------------------------------- 1 | function [net,f_measure,g_mean,precision,recall,err] = smv(net, data_train, labels_train, data_test, labels_test) 2 | % [net,f_measure,g_mean,precision,recall,err] = wmv(net, ... 3 | % data_train, labels_train, ... 4 | % data_test, labels_test, ... 5 | % smote_params) 6 | % 7 | % @net - initialized structure. you must initialize 8 | % net.mclass - number of classes 9 | % net.base_classifier - you should set this to be model.type 10 | % which is submitted to CLASSIFIER_TRAIN.m 11 | % net.n_classifiers - number of classifiers to keep in the pool 12 | % @data_train - cell array of training data. each entry should 13 | % have a n_oberservation by n_feature matrix 14 | % @labels_train - cell array of class labels 15 | % @data_test - cell array of training data. each entry should 16 | % have a n_oberservation by n_feature matrix 17 | % @labels_test - cell array of class labels 18 | % 19 | % Simple Majority Vote. 20 | % 21 | % @Author: Gregory Ditzler (gregory.ditzler@gmail.com) 22 | % 23 | % See also 24 | % CLASSIFIER_TRAIN.m CLASSIFIER_TEST.m 25 | 26 | % follow_the_leader.m 27 | % Copyright (C) 2013 Gregory Ditzler 28 | % 29 | % This program is free software: you can redistribute it and/or modify 30 | % it under the terms of the GNU General Public License as published by 31 | % the Free Software Foundation, either version 3 of the License, or 32 | % (at your option) any later version. 33 | % 34 | % This program is distributed in the hope that it will be useful, 35 | % but WITHOUT ANY WARRANTY; without even the implied warranty of 36 | % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 37 | % GNU General Public License for more details. 38 | % 39 | % You should have received a copy of the GNU General Public License 40 | % along with this program. If not, see . 41 | 42 | 43 | 44 | 45 | n_timestamps = length(data_train); % total number of time stamps 46 | net.classifiers = cell(net.n_classifiers,1); 47 | net.type = 'smv'; 48 | 49 | f_measure = zeros(n_timestamps, net.mclass); 50 | g_mean = zeros(n_timestamps, 1); 51 | recall = zeros(n_timestamps, net.mclass); 52 | precision = zeros(n_timestamps, net.mclass); 53 | err = zeros(n_timestamps, 1); 54 | 55 | if net.n_classifiers == -1 56 | net.n_classifiers = n_timestamps; 57 | end 58 | 59 | for ell = 1:n_timestamps 60 | % get the training data for the 't'th round 61 | data_train_t = data_train{ell}; 62 | labels_train_t = labels_train{ell}; 63 | data_test_t = data_test{ell}; 64 | labels_test_t = labels_test{ell}; 65 | index = mod(ell-1, net.n_classifiers) + 1; 66 | 67 | net.classifiers{index} = classifier_train(... 68 | net.base_classifier, ... 69 | data_train_t, ... 70 | labels_train_t); 71 | if ell < net.n_classifiers 72 | T = ell; 73 | else 74 | T = net.n_classifiers; 75 | end 76 | y = decision_ensemble(net, data_train_t, labels_train_t, T); 77 | e = zeros(T,1); 78 | for t = 1:T 79 | [~,~,~,~, e(t)] = stats(labels_train_t, y(:,t), net.mclass); 80 | end 81 | predictions = classify_ensemble(net.classifiers(1:T), ones(1,T)/T, ... 82 | net.mclass, data_test_t, labels_test_t); 83 | [f_measure(ell,:),g_mean(ell),recall(ell,:),precision(ell,:),... 84 | err(ell)] = stats(labels_test_t, predictions, net.mclass); 85 | end 86 | 87 | 88 | 89 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 90 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 91 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 92 | %%% AUXILARY FUNCTIONS 93 | function y = decision_ensemble(net, data, labels, n_experts) 94 | y = zeros(numel(labels), n_experts); 95 | for k = 1:n_experts 96 | y(:, k) = classifier_test(net.classifiers{k}, data); 97 | end 98 | 99 | 100 | function [predictions,posterior] = classify_ensemble(classifiers, weights, ... 101 | mclass, data, labels) 102 | n_experts = length(classifiers); 103 | if n_experts ~= length(weights) 104 | error('What are there are different number of weights and experts!') 105 | end 106 | p = zeros(numel(labels), mclass); 107 | for k = 1:n_experts 108 | y = classifier_test(classifiers{k}, data); 109 | 110 | % this is inefficient, but it does the job 111 | for m = 1:numel(y) 112 | p(m,y(m)) = p(m,y(m)) + weights(k); 113 | end 114 | end 115 | [~,predictions] = max(p'); 116 | predictions = predictions'; 117 | posterior = p./repmat(sum(p,2),1,mclass); 118 | 119 | -------------------------------------------------------------------------------- /src/stats.m: -------------------------------------------------------------------------------- 1 | function [f_measure,g_mean,recall,precision,err] = stats(f, h, mclass) 2 | % [f_measure,g_mean,recall,precision,err] = stats(f, h, mclass) 3 | % @f - vector of true labels 4 | % @h - vector of predictions on f 5 | % @mclass - number of classes 6 | % @f_measure 7 | % @g_mean 8 | % @recall 9 | % @precision 10 | % @err 11 | % 12 | 13 | % stats.m 14 | % Copyright (C) 2013 Gregory Ditzler 15 | % 16 | % This program is free software: you can redistribute it and/or modify 17 | % it under the terms of the GNU General Public License as published by 18 | % the Free Software Foundation, either version 3 of the License, or 19 | % (at your option) any later version. 20 | % 21 | % This program is distributed in the hope that it will be useful, 22 | % but WITHOUT ANY WARRANTY; without even the implied warranty of 23 | % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 24 | % GNU General Public License for more details. 25 | % 26 | % You should have received a copy of the GNU General Public License 27 | % along with this program. If not, see . 28 | 29 | F = index2vector(f, mclass); 30 | H = index2vector(h, mclass); 31 | 32 | recall = compute_recall(F, H, mclass); 33 | err = 1 - sum(diag(H'*F))/sum(sum(H'*F)); 34 | precision = compute_precision(F, H, mclass); 35 | g_mean = compute_g_mean(recall, mclass); 36 | f_measure = compute_f_measure(F, H, mclass); 37 | 38 | function g_mean = compute_g_mean(recall, mclass) 39 | g_mean = (prod(recall))^(1/mclass); 40 | 41 | function f_measure = compute_f_measure(F, H, mclass) 42 | f_measure = zeros(1, mclass); 43 | for c = 1:mclass 44 | f_measure(c) = 2*F(:, c)'*H(:, c)/(sum(H(:, c)) + sum(F(:, c))); 45 | end 46 | f_measure(isnan(f_measure)) = 1; 47 | 48 | function precision = compute_precision(F, H, mclass) 49 | precision = zeros(1, mclass); 50 | for c = 1:mclass 51 | precision(c) = F(:, c)'*H(:, c)/sum(H(:, c)); 52 | end 53 | precision(isnan(precision)) = 1; 54 | 55 | function recall = compute_recall(F, H, mclass) 56 | recall = zeros(1, mclass); 57 | for c = 1:mclass 58 | recall(c) = F(:, c)'*H(:, c)/sum(F(:, c)); 59 | end 60 | recall(isnan(recall)) = 1; 61 | 62 | function y = index2vector(x, mclass) 63 | y = zeros(numel(x), mclass); 64 | for n = 1:numel(x) 65 | y(n, x(n)) = 1; 66 | end 67 | -------------------------------------------------------------------------------- /src/swmv.m: -------------------------------------------------------------------------------- 1 | function [net,f_measure,g_mean,precision,recall,err] = swmv(net, data_train, labels_train, data_test, labels_test) 2 | % [net,f_measure,g_mean,precision,recall,err] = wmv(net, ... 3 | % data_train, labels_train, ... 4 | % data_test, labels_test, ... 5 | % smote_params) 6 | % 7 | % @net - initialized structure. you must initialize 8 | % net.mclass - number of classes 9 | % net.base_classifier - you should set this to be model.type 10 | % which is submitted to CLASSIFIER_TRAIN.m 11 | % net.n_classifiers - number of classifiers to keep in the pool 12 | % @data_train - cell array of training data. each entry should 13 | % have a n_oberservation by n_feature matrix 14 | % @labels_train - cell array of class labels 15 | % @data_test - cell array of training data. each entry should 16 | % have a n_oberservation by n_feature matrix 17 | % @labels_test - cell array of class labels 18 | % 19 | % Simple Weighted Majority Vote. 20 | % 21 | % @Author: Gregory Ditzler (gregory.ditzler@gmail.com) 22 | % 23 | % See also 24 | % CLASSIFIER_TRAIN.m CLASSIFIER_TEST.m 25 | 26 | % follow_the_leader.m 27 | % Copyright (C) 2013 Gregory Ditzler 28 | % 29 | % This program is free software: you can redistribute it and/or modify 30 | % it under the terms of the GNU General Public License as published by 31 | % the Free Software Foundation, either version 3 of the License, or 32 | % (at your option) any later version. 33 | % 34 | % This program is distributed in the hope that it will be useful, 35 | % but WITHOUT ANY WARRANTY; without even the implied warranty of 36 | % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 37 | % GNU General Public License for more details. 38 | % 39 | % You should have received a copy of the GNU General Public License 40 | % along with this program. If not, see . 41 | 42 | 43 | 44 | 45 | n_timestamps = length(data_train); % total number of time stamps 46 | net.classifiers = cell(net.n_classifiers,1); 47 | net.weights = zeros(net.n_classifiers,1); 48 | net.type = 'swmv'; 49 | 50 | f_measure = zeros(n_timestamps, net.mclass); 51 | g_mean = zeros(n_timestamps, 1); 52 | recall = zeros(n_timestamps, net.mclass); 53 | precision = zeros(n_timestamps, net.mclass); 54 | err = zeros(n_timestamps, 1); 55 | 56 | if net.n_classifiers == -1 57 | net.n_classifiers = n_timestamps; 58 | end 59 | 60 | for ell = 1:n_timestamps 61 | % get the training data for the 't'th round 62 | data_train_t = data_train{ell}; 63 | labels_train_t = labels_train{ell}; 64 | data_test_t = data_test{ell}; 65 | labels_test_t = labels_test{ell}; 66 | index = mod(ell-1, net.n_classifiers) + 1; 67 | 68 | net.classifiers{index} = classifier_train(... 69 | net.base_classifier, ... 70 | data_train_t, ... 71 | labels_train_t); 72 | if ell < net.n_classifiers 73 | T = ell; 74 | else 75 | T = net.n_classifiers; 76 | end 77 | y = decision_ensemble(net, data_train_t, labels_train_t, T); 78 | e = zeros(T,1); 79 | for t = 1:T 80 | [~,~,~,~, e(t)] = stats(labels_train_t, y(:,t), net.mclass); 81 | net.weights(t) = log((1 - e(t) + sqrt(eps))/(e(t) + sqrt(eps))); 82 | end 83 | predictions = classify_ensemble(net.classifiers(1:T), net.weights(1:T), ... 84 | net.mclass, data_test_t, labels_test_t); 85 | [f_measure(ell,:),g_mean(ell),recall(ell,:),precision(ell,:),... 86 | err(ell)] = stats(labels_test_t, predictions, net.mclass); 87 | end 88 | 89 | 90 | 91 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 92 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 93 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 94 | %%% AUXILARY FUNCTIONS 95 | function y = decision_ensemble(net, data, labels, n_experts) 96 | y = zeros(numel(labels), n_experts); 97 | for k = 1:n_experts 98 | y(:, k) = classifier_test(net.classifiers{k}, data); 99 | end 100 | 101 | 102 | function [predictions,posterior] = classify_ensemble(classifiers, weights, ... 103 | mclass, data, labels) 104 | n_experts = length(classifiers); 105 | if n_experts ~= length(weights) 106 | error('Why are there are different number of weights and experts!') 107 | end 108 | p = zeros(numel(labels), mclass); 109 | for k = 1:n_experts 110 | y = classifier_test(classifiers{k}, data); 111 | 112 | % this is inefficient, but it does the job 113 | for m = 1:numel(y) 114 | p(m,y(m)) = p(m,y(m)) + weights(k); 115 | end 116 | end 117 | [~,predictions] = max(p'); 118 | predictions = predictions'; 119 | posterior = p./repmat(sum(p,2),1,mclass); 120 | 121 | -------------------------------------------------------------------------------- /src/test_then_train.m: -------------------------------------------------------------------------------- 1 | function [data_train,data_test,labels_train,labels_test] = test_then_train(data, labels, win_size) 2 | % [data_train,data_test,labels_train,labels_test] = ... 3 | % test_then_train(data, labels, win_size); 4 | % 5 | % @data - data in n_observations by n_features matrix 6 | % @labels - labels in n_observations by 1 vector 7 | % @win_size - batch size 8 | % @data_train - cell array of training data 9 | % @data_test - cell array of test data 10 | % @labels_train - cell array of training labels 11 | % @labels_test - cell array of test labels 12 | % 13 | % 14 | % Partition a data set in fixed length windows for training and 15 | % testing. 16 | % 17 | % @Author: Gregory Ditzler (gregory.ditzler@gmail.com) 18 | % 19 | 20 | % test_then_train.m 21 | % Copyright (C) 2013 Gregory Ditzler 22 | % 23 | % This program is free software: you can redistribute it and/or modify 24 | % it under the terms of the GNU General Public License as published by 25 | % the Free Software Foundation, either version 3 of the License, or 26 | % (at your option) any later version. 27 | % 28 | % This program is distributed in the hope that it will be useful, 29 | % but WITHOUT ANY WARRANTY; without even the implied warranty of 30 | % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 31 | % GNU General Public License for more details. 32 | % 33 | % You should have received a copy of the GNU General Public License 34 | % along with this program. If not, see . 35 | n_observations = length(labels); 36 | n = 0; 37 | kill_loop = false; 38 | data_train = {}; 39 | data_test = {}; 40 | labels_train = {}; 41 | labels_test = {}; 42 | 43 | while true 44 | te_idx = (n+1)*win_size+1:(n+2)*win_size; 45 | tr_idx = n*win_size+1:(n+1)*win_size; 46 | n = n + 1; 47 | if max(te_idx) > n_observations 48 | te_idx = te_idx(te_idx <= n_observations); 49 | kill_loop = true; 50 | end 51 | 52 | data_train{n} = data(tr_idx, :); 53 | data_test{n} = data(te_idx, :); 54 | labels_train{n} = labels(tr_idx); 55 | labels_test{n} = labels(te_idx); 56 | 57 | if kill_loop == true 58 | break; 59 | end 60 | end -------------------------------------------------------------------------------- /tests/README.md: -------------------------------------------------------------------------------- 1 | IncrementalLearning 2 | =================== 3 | 4 | I am planning on providing some implementations of the Learn++ algorithm in Matlab. Please direct any questions, comments, or suggestions to . More details to come. 5 | -------------------------------------------------------------------------------- /tests/test_ftl.m: -------------------------------------------------------------------------------- 1 | clc; 2 | clear; 3 | close all; 4 | 5 | 6 | disp('The ConceptDriftData.m file must be in the Matlab path. This'); 7 | disp('file can be found: https://github.com/gditzler/ConceptDriftData '); 8 | addpath('../src/'); 9 | 10 | model.type = 'CART'; % base classifier 11 | net.mclass = 2; % number of classes in the prediciton problem 12 | net.base_classifier = model; % set the base classifier in the net struct 13 | net.n_classifiers = 10; 14 | 15 | % generate the sea data set 16 | T = 200; % number of time stamps 17 | N = 100; % number of data points at each time 18 | [data_train, labels_train,data_test,labels_test] = ConceptDriftData('sea', T, N); 19 | for t = 1:T 20 | % i wrote the code along time ago and i used at assume column vectors for 21 | % data and i wrote all the code for learn++ on github to assume row 22 | % vectors. the primary reasoning for this is that the stats toolbox in 23 | % matlab uses row vectors for operations like mean, cov and the 24 | % classifiers like CART and NB 25 | data_train{t} = data_train{t}'; 26 | labels_train{t} = labels_train{t}'; 27 | data_test{t} = data_test{t}'; 28 | labels_test{t} = labels_test{t}'; 29 | end 30 | 31 | % run learn++.nse 32 | [net,f_measure,g_mean,precision,recall,err] = follow_the_leader(net, data_train, labels_train, data_test, ... 33 | labels_test); 34 | 35 | plot(err) -------------------------------------------------------------------------------- /tests/test_learn.m: -------------------------------------------------------------------------------- 1 | function test_learn() 2 | % test learn++ 3 | K = 5; 4 | 5 | addpath('../src/'); % add the code path 6 | load ionosphere % load the built in ionosphere data set 7 | u = unique(Y); % get the number of unique classes 8 | labels = zeros(numel(Y), 1); 9 | 10 | % convert the string labels to numeric labels 11 | for n = 1:numel(Y) 12 | for c = 1:numel(u) 13 | if u{c} == Y{n} 14 | labels(n) = c; 15 | break 16 | end 17 | end 18 | end 19 | 20 | % shuffle the data 21 | i = randperm(numel(Y)); 22 | data = X(i, : ); 23 | labels = labels(i); 24 | clear Description X Y c i n u 25 | 26 | cv = cvpartition(numel(labels),'k',K); 27 | z = zeros(numel(labels),1); 28 | for k = 1:K-1 29 | z = z + (training(cv,k)>0); 30 | end 31 | ts_idx = find(z == K - 1); 32 | tr_idx = find(z ~= K - 1); 33 | 34 | 35 | data_tr = data(tr_idx, :); 36 | data_te = data(ts_idx, :); 37 | labels_tr = labels(tr_idx); 38 | labels_te = labels(ts_idx); 39 | 40 | cv = cvpartition(numel(labels_tr),'k',K); 41 | for k = 1:K 42 | data_tr_cell{k} = data_tr(training(cv,k)==0, :); 43 | labels_tr_cell{k} = labels_tr(training(cv,k)==0); 44 | end 45 | clear K cv data labels z tr_idx ts_idx k data_tr labels_tr 46 | 47 | model.type = 'CART'; 48 | net.base_classifier = model; 49 | net.iterations = 3; 50 | net.mclass = numel(unique(labels_te)); 51 | 52 | [net,errs] = learn(net, data_tr_cell, labels_tr_cell, data_te, labels_te); 53 | plot(errs) -------------------------------------------------------------------------------- /tests/test_learn_nse.m: -------------------------------------------------------------------------------- 1 | function test_learn_nse() 2 | % test learn++.nse 3 | 4 | disp('The ConceptDriftData.m file must be in the Matlab path. This'); 5 | disp('file can be found: https://github.com/gditzler/ConceptDriftData '); 6 | addpath('../src/'); 7 | 8 | model.type = 'CART'; % base classifier 9 | net.a = .5; % slope parameter to a sigmoid 10 | net.b = 10; % cutoff parameter to a sigmoid 11 | net.threshold = 0.01; % how small is too small for error 12 | net.mclass = 2; % number of classes in the prediciton problem 13 | net.base_classifier = model; % set the base classifier in the net struct 14 | 15 | % generate the sea data set 16 | T = 200; % number of time stamps 17 | N = 100; % number of data points at each time 18 | [data_train, labels_train,data_test,labels_test] = ConceptDriftData('sea', T, N); 19 | for t = 1:T 20 | % i wrote the code along time ago and i used at assume column vectors for 21 | % data and i wrote all the code for learn++ on github to assume row 22 | % vectors. the primary reasoning for this is that the stats toolbox in 23 | % matlab uses row vectors for operations like mean, cov and the 24 | % classifiers like CART and NB 25 | data_train{t} = data_train{t}'; 26 | labels_train{t} = labels_train{t}'; 27 | data_test{t} = data_test{t}'; 28 | labels_test{t} = labels_test{t}'; 29 | end 30 | 31 | % run learn++.nse 32 | [~,~,~,~,~,errs_nse] = learn_nse(net, data_train, labels_train, data_test, ... 33 | labels_test); 34 | 35 | % reset the parameters of the net struct. 36 | model.type = 'CART'; 37 | net.a = .5; 38 | net.b = 10; 39 | net.threshold = 0.01; 40 | net.mclass = 2; 41 | net.base_classifier = model; 42 | 43 | % set the parameters for smote 44 | smote_params.minority_class = 2; 45 | smote_params.k = 3; 46 | smote_params.N = 200; 47 | 48 | % run learn++.cds. the difference between calling cds or nse is that the 49 | % you pass the smote structure into the learn_nse.m function. 50 | [~,~,~,~,~,errs_cds] = learn_nse(net, data_train, labels_train, data_test, ... 51 | labels_test, smote_params); 52 | 53 | figure; 54 | plot(errs_nse) 55 | plot(errs_cds,'r') -------------------------------------------------------------------------------- /tests/test_smv.m: -------------------------------------------------------------------------------- 1 | clc; 2 | clear; 3 | close all; 4 | 5 | 6 | disp('The ConceptDriftData.m file must be in the Matlab path. This'); 7 | disp('file can be found: https://github.com/gditzler/ConceptDriftData '); 8 | addpath('../src/'); 9 | 10 | model.type = 'CART'; % base classifier 11 | net.mclass = 2; % number of classes in the prediciton problem 12 | net.base_classifier = model; % set the base classifier in the net struct 13 | net.n_classifiers = 10; 14 | 15 | % generate the sea data set 16 | T = 200; % number of time stamps 17 | N = 100; % number of data points at each time 18 | [data_train, labels_train,data_test,labels_test] = ConceptDriftData('sea', T, N); 19 | for t = 1:T 20 | % i wrote the code along time ago and i used at assume column vectors for 21 | % data and i wrote all the code for learn++ on github to assume row 22 | % vectors. the primary reasoning for this is that the stats toolbox in 23 | % matlab uses row vectors for operations like mean, cov and the 24 | % classifiers like CART and NB 25 | data_train{t} = data_train{t}'; 26 | labels_train{t} = labels_train{t}'; 27 | data_test{t} = data_test{t}'; 28 | labels_test{t} = labels_test{t}'; 29 | end 30 | 31 | [net,f_measure,g_mean,precision,recall,err] = smv(net, data_train, labels_train, data_test, ... 32 | labels_test); 33 | 34 | plot(err) -------------------------------------------------------------------------------- /tests/test_swmv.m: -------------------------------------------------------------------------------- 1 | clc; 2 | clear; 3 | close all; 4 | 5 | 6 | disp('The ConceptDriftData.m file must be in the Matlab path. This'); 7 | disp('file can be found: https://github.com/gditzler/ConceptDriftData '); 8 | addpath('../src/'); 9 | 10 | model.type = 'CART'; % base classifier 11 | net.mclass = 2; % number of classes in the prediciton problem 12 | net.base_classifier = model; % set the base classifier in the net struct 13 | net.n_classifiers = 10; 14 | 15 | % generate the sea data set 16 | T = 200; % number of time stamps 17 | N = 100; % number of data points at each time 18 | [data_train, labels_train,data_test,labels_test] = ConceptDriftData('sea', T, N); 19 | for t = 1:T 20 | % i wrote the code along time ago and i used at assume column vectors for 21 | % data and i wrote all the code for learn++ on github to assume row 22 | % vectors. the primary reasoning for this is that the stats toolbox in 23 | % matlab uses row vectors for operations like mean, cov and the 24 | % classifiers like CART and NB 25 | data_train{t} = data_train{t}'; 26 | labels_train{t} = labels_train{t}'; 27 | data_test{t} = data_test{t}'; 28 | labels_test{t} = labels_test{t}'; 29 | end 30 | 31 | [net,f_measure,g_mean,precision,recall,err] = swmv(net, data_train, labels_train, data_test, ... 32 | labels_test); 33 | 34 | plot(err) -------------------------------------------------------------------------------- /tests/test_ttt.m: -------------------------------------------------------------------------------- 1 | clc; 2 | clear; 3 | close all; 4 | 5 | addpath('../src/'); 6 | 7 | n_observations = 1000; 8 | n_features = 10; 9 | window = 100; 10 | data = randn(n_observations,n_features); 11 | labels = randi(2,1,n_observations)'; 12 | [x1,x2,y1,y2] = test_then_train(data, labels, window); --------------------------------------------------------------------------------