├── .gitignore ├── CONTRIBUTING.md ├── Example Notebook └── Auto_ViML_Demo.ipynb ├── LICENSE ├── PKG-INFO ├── README.md ├── SULOV.jpg ├── autoviml ├── Auto_NLP.py ├── Auto_ViML.py ├── QuickML_Ensembling.py ├── QuickML_Stacking.py ├── Transform_KM_Features.py ├── __init__.py ├── __version__.py ├── classify_method.py ├── custom_scores.py ├── custom_scores_HO.py ├── feature_engineering.py └── sulov_method.py ├── logo.png ├── requirements.txt ├── setup.cfg ├── setup.py └── sulov_xgboost.png /.gitignore: -------------------------------------------------------------------------------- 1 | .ipynb_checkpoints/ 2 | __pycache__/ 3 | .idea/ 4 | dist/ 5 | autoviz.egg-info/ 6 | autoviml.egg-info/ 7 | build/ 8 | diagnosis/ -------------------------------------------------------------------------------- /CONTRIBUTING.md: -------------------------------------------------------------------------------- 1 | # Contributing 2 | 3 | We welcome contributions from anyone beginner or advanced. Please before working on some feature 4 | 5 | * search through the past issues, your concern may have been raised by others in the past. Check through 6 | closed issues as well. 7 | * if there is no open issue for your feature request please open one up to coordinate all collaborators 8 | * write your feature 9 | * submit a pull request on this repo with: 10 | * a brief description 11 | * **detail of the expected change(s) in behaviour** 12 | * how to test it (if it's not obvious) 13 | 14 | Ask someone to test it. 15 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /PKG-INFO: -------------------------------------------------------------------------------- 1 | Metadata-Version: 2.1 2 | Name: autoviml 3 | Version: 0.1.669 4 | Summary: Automatically Build Variant Interpretable ML models fast - now with CatBoost! 5 | Home-page: https://github.com/AutoViML/Auto_ViML 6 | Author: Ram Seshadri 7 | License: Apache License 2.0 8 | Description: # Auto-ViML 9 | 10 | ![banner](logo.png) 11 | 12 | [![Downloads](https://pepy.tech/badge/autoviml/week)](https://pepy.tech/project/autoviml/week) 13 | [![Downloads](https://pepy.tech/badge/autoviml/month)](https://pepy.tech/project/autoviml/month) 14 | [![Downloads](https://pepy.tech/badge/autoviml)](https://pepy.tech/project/autoviml) 15 | [![standard-readme compliant](https://img.shields.io/badge/standard--readme-OK-green.svg?style=flat-square)](https://github.com/RichardLitt/standard-readme) 16 | [![Python Versions](https://img.shields.io/pypi/pyversions/autoviml.svg?logo=python&logoColor=white)](https://pypi.org/project/autoviml) 17 | [![PyPI Version](https://img.shields.io/pypi/v/autoviml.svg?logo=pypi&logoColor=white)](https://pypi.org/project/autoviml) 18 | [![PyPI License](https://img.shields.io/pypi/l/autoviml.svg)](https://github.com/AutoViML/Auto_ViML/blob/master/LICENSE) 19 | 20 | Automatically Build Variant Interpretable ML models fast! 21 | Auto_ViML is pronounced "auto vimal" (autovimal logo created by Sanket Ghanmare) 22 |

NEW FEATURES in this version are:
23 | 1. SMOTE -> now we use SMOTE for imbalanced data. Just set Imbalanced_Flag = True in input below
24 | 2. Auto_NLP: It automatically detects Text variables and does NLP processing on those columns
25 | 3. Date Time Variables: It automatically detects date time variables and adds extra features
26 | 4. Feature Engineering: Now you can perform feature engineering with the available featuretools library.
27 |

To upgrade to the best, most stable and full-featured version (anything over > 0.1.600), do one of the following:
28 | Use $ pip install autoviml --upgrade --ignore-installed
29 | or 30 | pip install git+https://github.com/AutoViML/Auto_ViML.git
31 | 32 | ## Table of Contents 33 |

43 | 44 | ## Background 45 | Read this [Medium article to learn how to use Auto_ViML](https://towardsdatascience.com/why-automl-is-an-essential-new-tool-for-data-scientists-2d9ab4e25e46).
46 |

Auto_ViML was designed for building High Performance Interpretable Models with the fewest variables. 47 | The "V" in Auto_ViML stands for Variable because it tries multiple models with multiple features to find you the best performing model for your dataset. The "i" in Auto_ViML stands for "interpretable" since Auto_ViML selects the least number of features necessary to build a simpler, more interpretable model. In most cases, Auto_ViML builds models with 20-99% fewer features than a similar performing model with all included features (this is based on my trials. Your experience may vary).
48 |

49 | Auto_ViML is every Data Scientist's model assistant that:

    50 |
  1. Helps you with data cleaning: you can send in your entire dataframe as is and Auto_ViML will suggest changes to help with missing values, formatting variables, adding variables, etc. It loves dirty data. The dirtier the better!
    51 |
  2. Assists you with variable classification: Auto_ViML classifies variables automatically. This is very helpful when you have hundreds if not thousands of variables since it can readily identify which of those are numeric vs categorical vs NLP text vs date-time variables and so on.
    52 |
  3. Performs feature reduction automatically. When you have small data sets and you know your domain well, it is easy to perhaps do EDA and identify which variables are important. But when you have a very large data set with hundreds if not thousands of variables, selecting the best features from your model can mean the difference between a bloated and highly complex model or a simple model with the fewest and most information-rich features. Auto_ViML uses XGBoost repeatedly to perform feature selection. You must try it on your large data sets and compare!
    53 |
  4. Produces model performance results as graphs automatically. Just set verbose = 1 (or) 2
    54 |
  5. Handles text, date-time, structs (lists, dictionaries), numeric, boolean, factor and categorical variables all in one model using one straight process.
    55 |
  6. Allows you to use the featuretools library to do Feature Engineering.
    See example below.
    56 | Let's say you have a few numeric features in your data called "preds". 57 | You can 'add','subtract','multiply' or 'divide' these features among themselves using this module. You can optionally send an ID column in the data so that the index ordering is preserved.
    58 |
    
     59 |         from autoviml.feature_engineering import feature_engineering
    60 | print(df[preds].shape)
    61 | dfmod = feature_engineering(df[preds],['add'],'ID')
    62 | print(dfmod.shape)
    63 |
    64 |
65 | Auto_ViML is built using scikit-learn, Nnumpy, pandas and matplotlib. It should run 66 | on most Python 3 Anaconda installations. You won't have to import any special 67 | libraries other than "XGBoost", "Imbalanced-Learn", "CatBoost", and "featuretools" library. We use "SHAP" library for interpretability.
But if you don't have these libraries, Auto_ViML will install those for you automatically. 68 | 69 | ## Install 70 | 71 | **Prerequsites:** 72 | 73 | - [Anaconda](https://docs.anaconda.com/anaconda/install/) 74 | 75 | To clone Auto_ViML, it is better to create a new environment, and install the required dependencies: 76 | 77 | To install from PyPi: 78 | 79 | ``` 80 | conda create -n python=3.7 anaconda 81 | conda activate # ON WINDOWS: `source activate ` 82 | pip install autoviml 83 | or 84 | pip install git+https://github.com/AutoViML/Auto_ViML.git 85 | ``` 86 | 87 | To install from source: 88 | 89 | ``` 90 | cd 91 | git clone git@github.com:AutoViML/Auto_ViML.git 92 | # or download and unzip https://github.com/AutoViML/Auto_ViML/archive/master.zip 93 | conda create -n python=3.7 anaconda 94 | conda activate # ON WINDOWS: `source activate ` 95 | cd Auto_ViML 96 | pip install -r requirements.txt 97 | ``` 98 | 99 | ## Usage 100 | 101 | In the same directory, open a Jupyter Notebook and use this line to import the .py file: 102 | 103 | ``` 104 | from autoviml.Auto_ViML import Auto_ViML 105 | ``` 106 | 107 | Load a data set (any CSV or text file) into a Pandas dataframe and split it into Train and Test dataframes. If you don't have a test dataframe, you can simple assign the test variable below to '' (empty string): 108 | 109 | ``` 110 | model, features, trainm, testm = Auto_ViML( 111 | train, 112 | target, 113 | test, 114 | sample_submission, 115 | hyper_param="GS", 116 | feature_reduction=True, 117 | scoring_parameter="weighted-f1", 118 | KMeans_Featurizer=False, 119 | Boosting_Flag=False, 120 | Binning_Flag=False, 121 | Add_Poly=False, 122 | Stacking_Flag=False, 123 | Imbalanced_Flag=False, 124 | verbose=0, 125 | ) 126 | ``` 127 | 128 | Finally, it writes your submission file to disk in the current directory called `mysubmission.csv`. 129 | This submission file is ready for you to show it clients or submit it to competitions. 130 | If no submission file was given, but as long as you give it a test file name, it will create a submission file for you named `mySubmission.csv`. 131 | Auto_ViML works on any Multi-Class, Multi-Label Data Set. So you can have many target labels. 132 | You don't have to tell Auto_ViML whether it is a Regression or Classification problem. 133 | 134 | ## Tips for using Auto_ViML: 135 | 1. For Classification problems and imbalanced classes, choose scoring_parameter="balanced_accuracy". It works better. 136 | 2. For Imbalanced Classes (<5% samples in rare class), choose "Imbalanced_Flag"=True. You can also set this flag to True for Regression problems where the target variable might have skewed distributions. 137 | 3. For Multi-Label dataset, the target input target variable can be sent in as a list of variables. 138 | 4. It is recommended that you first set Boosting_Flag=None to get a Linear model. Once you understand that, then you can try to set Boosting_Flag=False to get a Random Forest model. Finally, try Boosting_Flag=True to get an XGBoost model. This is the order that we recommend in order to use Auto_ViML. 139 | 5. Finally try Boosting_Flag="CatBoost" to get a complex but high performing model. 140 | 6. Binning_Flag=True improves a CatBoost model since it adds to the list of categorical vars in data 141 | 7. KMeans_featurizer=True works well in NLP and CatBoost models since it creates cluster variables 142 | 8. Add_Poly=3 improves certain models where there is date-time or categorical and numeric variables 143 | 9. feature_reduction=True is the default and works best. But when you have <10 features in data, set it to False 144 | 10. Do not use Stacking_Flag=True with Linear models since your results may not look great. 145 | 11. Use Stacking_Flag=True only for complex models and as a last step with Boosting_Flag=True or CatBoost 146 | 12. Always set hyper_param ="RS" as input since it runs faster than GridSearchCV and gives better results! 147 | 13. KMeans_Featurizer=True does not work well for small data sets. Use it for data sets > 10,000 rows. 148 | 14. Finally Auto_ViML is meant to be a baseline or challenger solution to your data set. So use it for making quick models that you can compare against or in Hackathons. It is not meant for production! 149 | 150 | ## API 151 | 152 | **Arguments** 153 | 154 | - `train`: could be a datapath+filename or a dataframe. It will detect which is which and load it. 155 | - `test`: could be a datapath+filename or a dataframe. If you don't have any, just leave it as "". 156 | - `submission`: must be a datapath+filename. If you don't have any, just leave it as empty string. 157 | - `target`: name of the target variable in the data set. 158 | - `sep`: if you have a spearator in the file such as "," or "\t" mention it here. Default is ",". 159 | - `scoring_parameter`: if you want your own scoring parameter such as "f1" give it here. If not, it will assume the appropriate scoring param for the problem and it will build the model. 160 | - `hyper_param`: Tuning options are GridSearch ('GS') and RandomizedSearch ('RS'). Default is 'RS'. 161 | - `feature_reduction`: Default = 'True' but it can be set to False if you don't want automatic feature_reduction since in Image data sets like digits and MNIST, you get better results when you don't reduce features automatically. You can always try both and see. 162 | - `KMeans_Featurizer` 163 | - `True`: Adds a cluster label to features based on KMeans. Use for Linear. 164 | - `False (default)` For Random Forests or XGB models, leave it False since it may overfit. 165 | - `Boosting Flag`: you have 4 possible choices (default is False): 166 | - `None` This will build a Linear Model 167 | - `False` This will build a Random Forest or Extra Trees model (also known as Bagging) 168 | - `True` This will build an XGBoost model 169 | - `CatBoost` This will build a CatBoost model (provided you have CatBoost installed) 170 | - `Add_Poly`: Default is 0 which means do-nothing. But it has three interesting settings: 171 | - `1` Add interaction variables only such as x1*x2, x2*x3,...x9\*10 etc. 172 | - `2` Add Interactions and Squared variables such as x1**2, x2**2, etc. 173 | - `3` Adds both Interactions and Squared variables such as x1*x2, x1**2,x2*x3, x2**2, etc. 174 | - `Stacking_Flag`: Default is False. If set to True, it will add an additional feature which is derived from predictions of another model. This is used in some cases but may result in overfitting. So be careful turning this flag "on". 175 | - `Binning_Flag`: Default is False. It set to True, it will convert the top numeric variables into binned variables through a technique known as "Entropy" binning. This is very helpful for certain datasets (especially hard to build models). 176 | - `Imbalanced_Flag`: Default is False. If set to True, it will use SMOTE from Imbalanced-Learn to oversample the "Rare Class" in an imbalanced dataset and make the classes balanced (50-50 for example in a binary classification). This also works for Regression problems where you have highly skewed distributions in the target variable. Auto_ViML creates additional samples using SMOTE for Highly Imbalanced data. 177 | - `verbose`: This has 3 possible states: 178 | - `0` limited output. Great for running this silently and getting fast results. 179 | - `1` more charts. Great for knowing how results were and making changes to flags in input. 180 | - `2` lots of charts and output. Great for reproducing what Auto_ViML does on your own. 181 | 182 | **Return values** 183 | 184 | - `model`: It will return your trained model 185 | - `features`: the fewest number of features in your model to make it perform well 186 | - `train_modified`: this is the modified train dataframe after removing and adding features 187 | - `test_modified`: this is the modified test dataframe with the same transformations as train 188 | 189 | ## Maintainers 190 | 191 | * [@AutoViML](https://github.com/AutoViML) 192 | * [@morenoh149](https://github.com/morenoh149) 193 | * [@hironroy](https://github.com/hironroy) 194 | 195 | ## Contributing 196 | 197 | See [the contributing file](CONTRIBUTING.md)! 198 | 199 | PRs accepted. 200 | 201 | ## License 202 | 203 | Apache License 2.0 © 2020 Ram Seshadri 204 | 205 | ## DISCLAIMER 206 | This project is not an official Google project. It is not supported by Google and Google specifically disclaims all warranties as to its quality, merchantability, or fitness for a particular purpose. 207 | 208 | Platform: UNKNOWN 209 | Classifier: Programming Language :: Python :: 3 210 | Classifier: Operating System :: OS Independent 211 | Description-Content-Type: text/markdown 212 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Auto-ViML 2 | 3 | ![banner](logo.png)
4 | Automatically Build Various Interpretable ML models fast!
5 | [![Downloads](https://pepy.tech/badge/autoviml/week)](https://pepy.tech/project/autoviml/week) 6 | [![Downloads](https://pepy.tech/badge/autoviml/month)](https://pepy.tech/project/autoviml/month) 7 | [![Downloads](https://pepy.tech/badge/autoviml)](https://pepy.tech/project/autoviml) 8 | [![standard-readme compliant](https://img.shields.io/badge/standard--readme-OK-green.svg?style=flat-square)](https://github.com/RichardLitt/standard-readme) 9 | [![Python Versions](https://img.shields.io/pypi/pyversions/autoviml.svg?logo=python&logoColor=white)](https://pypi.org/project/autoviml) 10 | [![PyPI Version](https://img.shields.io/pypi/v/autoviml.svg?logo=pypi&logoColor=white)](https://pypi.org/project/autoviml) 11 | [![PyPI License](https://img.shields.io/pypi/l/autoviml.svg)](https://github.com/AutoViML/Auto_ViML/blob/master/LICENSE) 12 | 13 | Auto_ViML is pronounced as "auto vimal" (autovimal logo created by Sanket Ghanmare). 14 | 15 | ## Update (Jan 2025) 16 |
    17 |
  1. Auto_ViML is now upgraded to version 0.2 which means it now runs on Python 3.12 or greater and also pandas 2.0 - this is a huge upgrade to those working in Colabs, Kaggle and other latest kernels. Please make sure you check the `requirements.txt` file to know which versions are recommended.
  2. 18 |
19 | 20 | ## Update (March 2023) 21 |
    22 |
  1. Auto_ViML has a new flag to speed up processing using GPU's. Just set the `GPU_flag`=`True` on Colab and other environments. But don't forget to set the runtime type to be "GPU" while running on Colab. Otherwise you will get an error.
  2. 23 |
24 | 25 | ## Update (May 2022) 26 |
    27 |
  • Auto_ViML as of version 0.1.710 uses a very high performance library called `imbalanced_ensemble` for Imbalanced dataset problems. It will produce a 5-10% boost in your balanced_accuracy based on my experience with many datasets I have tried.
  • 28 |
29 |
30 |

In addition, new features in this version are:
31 |

    32 |
  • SULOV -> Uses the SULOV algorithm for removing highly correlated features automatically.
  • 33 |
  • Auto_NLP -> AutoViML automatically detects Text variables and does NLP processing using Auto_NLP
  • 34 |
  • Date Time -> It automatically detects date time variables and generates new features
  • 35 |
  • `imbalanced_ensemble` library -> Uses imbalanced_ensemble library for imbalanced data. Just set Imbalanced_Flag = True in arguments
  • 36 |
  • Feature Selection -> We use the same algorithm that featurewiz library uses: SULOV and Recursive XGBoost to select best features fast. See below.
  • 37 |
38 | 39 | ## Table of Contents 40 | 50 | 51 | ## Background 52 |

Read this Medium article to learn how to use Auto_ViML effectively. 53 | 54 |

Auto_ViML was designed for building High Performance Interpretable Models with the fewest variables needed. 55 | The "V" in Auto_ViML stands for Variant because it tries multiple models with multiple features to find you the best performing model for your dataset. The "i" in Auto_ViML stands for "interpretable" since Auto_ViML selects the least number of features necessary to build a simpler, more interpretable model. In most cases, Auto_ViML builds models with 20%-99% fewer features than a similar performing model with all included features (this is based on my trials. Your experience may vary).
56 |

57 | Auto_ViML is every Data Scientist's model accelerator tool that: 58 |

    59 |
  1. Helps you with data cleaning: you can send in your entire dataframe as is and Auto_ViML will suggest changes to help with missing values, formatting variables, adding variables, etc. It loves dirty data. The dirtier the better!
  2. 60 |
  3. Performs Feature Selection: Auto_ViML selects variables automatically by default. This is very helpful when you have hundreds if not thousands of variables since it can readily identify which of those are important variables vs which are unnecessary. You can turn this off as well (see API).
  4. 61 | 62 | ![xgboost](sulov_xgboost.png) 63 |
  5. Removes highly correlated features automatically. If two variables are highly correlated in your dataset, which one should you remove and which one should you keep? The decision is not as easy as it looks. Auto_ViML uses the SULOV algorithm to remove highly correlated features. You can understand SULOV from this picture more intuitively.
  6. 64 | 65 | ![sulov](SULOV.jpg) 66 | 67 |
  7. Generates performance results graphically. Just set verbose = 1 (or) 2 instead of 0 (silent). You will get higher quality of insights as you increase verbosity.
  8. 68 |
  9. Handles text, date-time, structs (lists, dictionaries), numeric, boolean, factor and categorical variables all in one model using one straight process.
  10. 69 |
70 | Auto_ViML is built using scikit-learn, numpy, pandas and matplotlib. It should run on most Python 3 Anaconda installations. You won't have to import any special libraries other than "XGBoost", "Imbalanced-Learn", "CatBoost", and "featuretools" library. We use "SHAP" library for interpretability.
But if you don't have these libraries, Auto_ViML will install those for you automatically. 71 | 72 | ## Install 73 | 74 | **Prerequsites:** 75 | 76 | - [Anaconda](https://docs.anaconda.com/anaconda/install/) 77 | 78 | To clone Auto_ViML, it is better to create a new environment, and install the required dependencies: 79 | 80 | To install from PyPi: 81 |

82 | $ pip install autoviml --upgrade --ignore-installed
83 | 84 | 85 | ``` 86 | pip install git+https://github.com/AutoViML/Auto_ViML.git 87 | ``` 88 | 89 | To install from source: 90 | 91 | ``` 92 | cd 93 | git clone git@github.com:AutoViML/Auto_ViML.git 94 | # or download and unzip https://github.com/AutoViML/Auto_ViML/archive/master.zip 95 | conda create -n python=3.7 anaconda 96 | conda activate # ON WINDOWS: `source activate ` 97 | cd Auto_ViML 98 | pip install -r requirements.txt 99 | ``` 100 | 101 | ## Usage 102 | 103 | In the same directory, open a Jupyter Notebook and use this line to import the .py file: 104 | 105 | ``` 106 | from autoviml.Auto_ViML import Auto_ViML 107 | ``` 108 | 109 | Load a data set (any CSV or text file) into a Pandas dataframe and split it into Train and Test dataframes. If you don't have a test dataframe, you can simple assign the test variable below to '' (empty string): 110 | 111 | ``` 112 | model, features, trainm, testm = Auto_ViML( 113 | train, 114 | target, 115 | test, 116 | sample_submission, 117 | hyper_param="GS", 118 | feature_reduction=True, 119 | scoring_parameter="weighted-f1", 120 | KMeans_Featurizer=False, 121 | Boosting_Flag=False, 122 | Binning_Flag=False, 123 | Add_Poly=False, 124 | Stacking_Flag=False, 125 | Imbalanced_Flag=False, 126 | verbose=0, 127 | ) 128 | ``` 129 | 130 | Finally, it writes your submission file to disk in the current directory called `mysubmission.csv`. 131 | This submission file is ready for you to show it clients or submit it to competitions. 132 | If no submission file was given, but as long as you give it a test file name, it will create a submission file for you named `mySubmission.csv`. 133 | Auto_ViML works on any Multi-Class, Multi-Label Data Set. So you can have many target labels. 134 | You don't have to tell Auto_ViML whether it is a Regression or Classification problem. 135 | 136 | ## Tips for using Auto_ViML: 137 | 1. `scoring_parameter`: For Classification problems and imbalanced classes, choose scoring_parameter="balanced_accuracy". It works better. 138 | 2. `Imbalanced_Flag`: For Imbalanced Classes (<5% samples in rare class), choose "Imbalanced_Flag"=True. You can also set this flag to True for Regression problems where the target variable might have skewed distributions. 139 | 3. `target`: For Multi-Label dataset, the target input target variable can be sent in as a list of variables. 140 | 4. `Boosting_Flag`: It is recommended that you first set Boosting_Flag=None to get a Linear model. Once you understand that, then you can try to set Boosting_Flag=False to get a Random Forest model. Finally, try Boosting_Flag=True to get an XGBoost model. This is the order that we recommend in order to use Auto_ViML. Finally try Boosting_Flag="CatBoost" to get a complex but high performing model. 141 | 5. `Binning_Flag`: Binning_Flag=True improves a CatBoost model since it adds to the list of categorical vars in data 142 | 6. `KMeans_featurizer`: KMeans_featurizer=True works well in NLP and CatBoost models since it creates cluster variables 143 | 7. `Add_Poly`: Add_Poly=3 improves certain models where there is date-time or categorical and numeric variables 144 | 8. `feature_reduction`: feature_reduction=True is the default and works best. But when you have <10 features in data, set it to False 145 | 9. `Stacking_Flag`: Do not set Stacking_Flag=True with Linear models since your results may not look great. 146 | 10. `Stacking_Flag`: Use Stacking_Flag=True only for complex models and as a last step with Boosting_Flag=True or CatBoost 147 | 11. `hyper_param`: Leave hyper_param ="RS" as input since it runs faster than GridSearchCV and gives better results unless you have a small data set and can afford to spend time on hyper tuning. 148 | 12. `KMeans_Featurizer`: KMeans_Featurizer=True does not work well for small data sets. Use it for data sets > 10,000 rows. 149 | 13. `Final thoughts`: Finally Auto_ViML is meant to be a baseline or challenger solution to your data set. So use it for making quick models that you can compare against or in Hackathons. It is not meant for production! 150 | 151 | ## API 152 | 153 | **Arguments** 154 | 155 | - `train`: could be a datapath+filename or a dataframe. It will detect which is which and load it. 156 | - `test`: could be a datapath+filename or a dataframe. If you don't have any, just leave it as "". 157 | - `submission`: must be a datapath+filename. If you don't have any, just leave it as empty string. 158 | - `target`: name of the target variable in the data set. 159 | - `sep`: if you have a spearator in the file such as "," or "\t" mention it here. Default is ",". 160 | - `scoring_parameter`: if you want your own scoring parameter such as "f1" give it here. If not, it will assume the appropriate scoring param for the problem and it will build the model. 161 | - `hyper_param`: Tuning options are GridSearch ('GS') and RandomizedSearch ('RS'). Default is 'RS'. 162 | - `feature_reduction`: Default = 'True' but it can be set to False if you don't want automatic feature_reduction since in Image data sets like digits and MNIST, you get better results when you don't reduce features automatically. You can always try both and see. 163 | - `KMeans_Featurizer` 164 | - `True`: Adds a cluster label to features based on KMeans. Use for Linear. 165 | - `False (default)` For Random Forests or XGB models, leave it False since it may overfit. 166 | - `Boosting Flag`: you have 4 possible choices (default is False): 167 | - `None` This will build a Linear Model 168 | - `False` This will build a Random Forest or Extra Trees model (also known as Bagging) 169 | - `True` This will build an XGBoost model 170 | - `CatBoost` This will build a CatBoost model (provided you have CatBoost installed) 171 | - `Add_Poly`: Default is 0 which means do-nothing. But it has three interesting settings: 172 | - `1` Add interaction variables only such as x1*x2, x2*x3,...x9\*10 etc. 173 | - `2` Add Interactions and Squared variables such as x1**2, x2**2, etc. 174 | - `3` Adds both Interactions and Squared variables such as x1*x2, x1**2,x2*x3, x2**2, etc. 175 | - `Stacking_Flag`: Default is False. If set to True, it will add an additional feature which is derived from predictions of another model. This is used in some cases but may result in overfitting. So be careful turning this flag "on". 176 | - `Binning_Flag`: Default is False. It set to True, it will convert the top numeric variables into binned variables through a technique known as "Entropy" binning. This is very helpful for certain datasets (especially hard to build models). 177 | - `Imbalanced_Flag`: Default is False. Uses imbalanced_ensemble library for imbalanced data. Just set Imbalanced_Flag = True in arguments 178 | - `verbose`: This has 3 possible states: 179 | - `0` limited output. Great for running this silently and getting fast results. 180 | - `1` more charts. Great for knowing how results were and making changes to flags in input. 181 | - `2` lots of charts and output. Great for reproducing what Auto_ViML does on your own. 182 | 183 | **Return values** 184 | 185 | - `model`: It will return your trained model 186 | - `features`: the fewest number of features in your model to make it perform well 187 | - `train_modified`: this is the modified train dataframe after removing and adding features 188 | - `test_modified`: this is the modified test dataframe with the same transformations as train 189 | 190 | ## Maintainers 191 | 192 | * [@AutoViML](https://github.com/AutoViML) 193 | * [@morenoh149](https://github.com/morenoh149) 194 | * [@hironroy](https://github.com/hironroy) 195 | 196 | ## Contributing 197 | 198 | See [the contributing file](CONTRIBUTING.md)! 199 | 200 | PRs accepted. 201 | 202 | ## License 203 | 204 | Apache License 2.0 © 2020 Ram Seshadri 205 | 206 | ## DISCLAIMER 207 | This project is not an official Google project. It is not supported by Google and Google specifically disclaims all warranties as to its quality, merchantability, or fitness for a particular purpose. 208 | -------------------------------------------------------------------------------- /SULOV.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AutoViML/Auto_ViML/0761412417b6dac0683425c372fd7e4f8942d815/SULOV.jpg -------------------------------------------------------------------------------- /autoviml/QuickML_Ensembling.py: -------------------------------------------------------------------------------- 1 | import matplotlib.pyplot as plt 2 | import numpy as np 3 | import pandas as pd 4 | import seaborn as sns 5 | 6 | sns.set(style="white", color_codes=True) 7 | import warnings 8 | 9 | warnings.filterwarnings("ignore") 10 | from sklearn.linear_model import SGDClassifier 11 | from sklearn.model_selection import StratifiedKFold, KFold 12 | from sklearn.discriminant_analysis import LinearDiscriminantAnalysis 13 | from sklearn.ensemble import BaggingRegressor, RandomForestClassifier, RandomForestRegressor 14 | from sklearn.svm import LinearSVC, LinearSVR 15 | from sklearn.ensemble import AdaBoostClassifier, AdaBoostRegressor 16 | from sklearn.tree import DecisionTreeRegressor 17 | from sklearn.linear_model import LassoLarsCV 18 | from sklearn.linear_model import LogisticRegressionCV 19 | from sklearn.naive_bayes import GaussianNB, MultinomialNB 20 | from sklearn.tree import DecisionTreeClassifier 21 | from sklearn.multiclass import OneVsRestClassifier 22 | from sklearn.neighbors import KNeighborsRegressor 23 | import time 24 | import operator 25 | 26 | 27 | ######################################################### 28 | def QuickML_Ensembling(X_train, y_train, X_test, y_test='', modeltype='Regression', 29 | Boosting_Flag=False, 30 | scoring='', verbose=0): 31 | """ 32 | Quickly builds and runs multiple models for a clean data set(only numerics). 33 | """ 34 | start_time = time.time() 35 | seed = 99 36 | FOLDS = 5 37 | model_dict = {} 38 | model_tuples = [] 39 | if len(X_train) <= 100000 and X_train.shape[1] < 50: 40 | NUMS = 100 41 | else: 42 | try: 43 | X_train = X_train.sample(frac=0.30, random_state=99) 44 | y_train = y_train[X_train.index] 45 | except: 46 | pass 47 | NUMS = 200 48 | ###### This is where we start performing ensembling of multiple simpler models ### 49 | if modeltype == 'Regression': 50 | if scoring == '': 51 | scoring = 'neg_mean_squared_error' 52 | # scv = ShuffleSplit(n_splits=FOLDS,random_state=seed) 53 | scv = KFold(n_splits=FOLDS, shuffle=False) 54 | if Boosting_Flag is None: 55 | ## Create an ensemble model #### 56 | model5 = AdaBoostRegressor(estimator=DecisionTreeRegressor( 57 | random_state=seed, max_depth=1, min_samples_leaf=2 58 | ), n_estimators=NUMS, random_state=seed) 59 | model_tuples.append(('Adaboost', model5)) 60 | elif not Boosting_Flag: 61 | model5 = LassoLarsCV(cv=scv) 62 | model_tuples.append(('LassoLarsCV', model5)) 63 | else: 64 | model5 = LassoLarsCV(cv=scv) 65 | model_tuples.append(('LassoLarsCV', model5)) 66 | if Boosting_Flag is None: 67 | model6 = BaggingRegressor(DecisionTreeRegressor(random_state=seed), 68 | n_estimators=NUMS, random_state=seed) 69 | model_tuples.append(('Bagging_Regressor', model6)) 70 | elif not Boosting_Flag: 71 | model6 = LinearSVR() 72 | model_tuples.append(('Linear_SVR', model6)) 73 | else: 74 | model6 = DecisionTreeRegressor(max_depth=5, min_samples_leaf=2) 75 | model_tuples.append(('Decision_Tree', model6)) 76 | model7 = KNeighborsRegressor(n_neighbors=8) 77 | model_tuples.append(('KNN_Regressor', model7)) 78 | if Boosting_Flag is None: 79 | #### If the Boosting_Flag is True, it means Boosting model is present. 80 | ### So choose a different kind of classifier here 81 | model8 = DecisionTreeRegressor(max_depth=5, min_samples_leaf=2) 82 | model_tuples.append(('Decision_Tree', model8)) 83 | elif not Boosting_Flag: 84 | #### If the Boosting_Flag is True, it means Boosting model is present. 85 | ### So choose a different kind of classifier here 86 | model8 = AdaBoostRegressor(base_estimator=DecisionTreeRegressor( 87 | random_state=seed, max_depth=1, min_samples_leaf=2 88 | ), n_estimators=NUMS, random_state=seed) 89 | model_tuples.append(('Adaboost', model8)) 90 | else: 91 | model8 = RandomForestRegressor(bootstrap=False, 92 | max_depth=10, 93 | max_features='auto', 94 | min_samples_leaf=2, 95 | n_estimators=200, 96 | random_state=99) 97 | model_tuples.append(('RF_Regressor', model8)) 98 | else: 99 | if scoring == '': 100 | scoring = 'accuracy' 101 | num_classes = len(np.unique(y_test)) 102 | scv = StratifiedKFold(n_splits=FOLDS, shuffle=True, random_state=seed) 103 | if Boosting_Flag is None: 104 | ## Create an ensemble model #### 105 | model5 = AdaBoostClassifier(base_estimator=DecisionTreeClassifier( 106 | random_state=seed, max_depth=1, min_samples_leaf=2 107 | ), n_estimators=NUMS, random_state=seed) 108 | model_tuples.append(('Adaboost', model5)) 109 | elif not Boosting_Flag: 110 | model5 = LinearDiscriminantAnalysis() 111 | model_tuples.append(('Linear_Discriminant', model5)) 112 | else: 113 | model5 = LogisticRegressionCV(Cs=[0.001, 0.01, 0.1, 1, 10, 100], tol=0.01, 114 | solver='liblinear', random_state=seed) 115 | model_tuples.append(('Logistic_Regression_CV', model5)) 116 | if Boosting_Flag is None: 117 | model6 = DecisionTreeClassifier(max_depth=5, min_samples_leaf=2) 118 | model_tuples.append(('Decision_Tree', model6)) 119 | elif not Boosting_Flag: 120 | model6 = LinearSVC() 121 | model_tuples.append(('Linear_SVC', model6)) 122 | else: 123 | model6 = DecisionTreeClassifier(max_depth=5, min_samples_leaf=2) 124 | model_tuples.append(('Decision_Tree', model6)) 125 | if modeltype == 'Binary_Classification': 126 | model7 = GaussianNB() 127 | else: 128 | model7 = MultinomialNB() 129 | model_tuples.append(('Naive_Bayes', model7)) 130 | if Boosting_Flag is None: 131 | #### If the Boosting_Flag is True, it means Boosting model is present. 132 | ### So choose a different kind of classifier here 133 | model8 = RandomForestClassifier(bootstrap=False, 134 | max_depth=10, 135 | max_features='auto', 136 | min_samples_leaf=2, 137 | n_estimators=200, 138 | random_state=99) 139 | model_tuples.append(('Bagging_Classifier', model8)) 140 | elif not Boosting_Flag: 141 | #### If the Boosting_Flag is True, it means Boosting model is present. 142 | ### So choose a different kind of classifier here 143 | sgd_best_model = SGDClassifier(alpha=1e-06, 144 | loss='log', 145 | max_iter=1000, 146 | penalty='l2', 147 | learning_rate='constant', 148 | eta0=.1, 149 | random_state=3, 150 | tol=None) 151 | model8 = OneVsRestClassifier(sgd_best_model) 152 | model_tuples.append(('One_vs_Rest_Classifier', model8)) 153 | else: 154 | model8 = RandomForestClassifier(bootstrap=False, 155 | max_depth=10, 156 | max_features='auto', 157 | min_samples_leaf=2, 158 | n_estimators=200, 159 | random_state=99) 160 | model_tuples.append(('Bagging_Classifier', model8)) 161 | model_dict = dict(model_tuples) 162 | models, results = run_ensemble_models(model_dict, X_train, y_train, X_test, y_test, 163 | scoring, modeltype) 164 | return models, results 165 | 166 | 167 | ######################################################### 168 | from sklearn.metrics import balanced_accuracy_score, accuracy_score, precision_score, recall_score, f1_score 169 | from sklearn.metrics import mean_absolute_error, mean_squared_error 170 | import copy 171 | from sklearn.linear_model import LinearRegression 172 | from sklearn.linear_model import LogisticRegression 173 | 174 | 175 | def run_ensemble_models(model_dict, X_train, y_train, X_test, y_test, scoring, modeltype): 176 | start_time = time.time() 177 | model_name, bac_score_list, ac_score_list, p_score_list, r_score_list, f1_score_list = [], [], [], [], [], [] 178 | iteration = 0 179 | estimators = [] 180 | estim_tuples = [] 181 | for key, val in model_dict.items(): 182 | estimator_name = key 183 | model_name.append(key) 184 | if str(val).split("(")[0] == 'MultinomialNB': 185 | #### Multinomial models need only positive values!! 186 | val.fit(abs(X_train), y_train) 187 | y_pred = val.predict(abs(X_test)) 188 | else: 189 | try: 190 | val.fit(X_train, y_train) 191 | y_pred = val.predict(X_test) 192 | except: 193 | if modeltype != 'Regression': 194 | key = 'Logistic_Regression' 195 | val = LogisticRegression() 196 | else: 197 | key = 'Linear_Regression' 198 | val = LinearRegression() 199 | val.fit(X_train, y_train) 200 | try: 201 | y_pred = val.predict(X_test) 202 | except: 203 | print('Error in ensemble models. Returning...') 204 | stacks = np.array([]) 205 | return model_name, stacks 206 | if iteration == 0: 207 | stacks = copy.deepcopy(y_pred) 208 | iteration += 1 209 | else: 210 | stacks = np.c_[stacks, y_pred] 211 | if not isinstance(y_test, str): 212 | if modeltype == 'Regression': 213 | bac_score = np.sqrt(mean_squared_error(y_test, y_pred)) 214 | estimators.append((estimator_name, val, bac_score)) 215 | estim_tuples.append((estimator_name, bac_score)) 216 | bac_score_list.append(bac_score) 217 | ac_score_list.append(mean_squared_error(y_test, y_pred)) 218 | p_score_list.append(mean_absolute_error(y_test, y_pred)) 219 | model_comparison_df = pd.DataFrame([model_name, bac_score_list, ac_score_list, p_score_list]).T 220 | model_comparison_df.columns = ['model_name', 'RMSE', 'MSE', 'MAE'] 221 | model_comparison_df = model_comparison_df.sort_values(by='RMSE', ascending=True) 222 | else: 223 | bac_score = balanced_accuracy_score(y_test, y_pred) 224 | estimators.append((estimator_name, val, bac_score)) 225 | estim_tuples.append((estimator_name, bac_score)) 226 | bac_score_list.append(balanced_accuracy_score(y_test, y_pred)) 227 | ac_score_list.append(accuracy_score(y_test, y_pred)) 228 | p_score_list.append(precision_score(y_test, y_pred, average='macro')) 229 | r_score_list.append(recall_score(y_test, y_pred, average='macro')) 230 | f1_score_list.append(f1_score(y_test, y_pred, average='macro')) 231 | model_comparison_df = pd.DataFrame([model_name, bac_score_list, ac_score_list, p_score_list, 232 | r_score_list, f1_score_list]).T 233 | model_comparison_df.columns = ['model_name', 'bal_accuracy_score', 'accuracy_score', 234 | 'ave_precision_score', 'ave_recall_score', 'ave_f1_score'] 235 | model_comparison_df = model_comparison_df.sort_values(by='bal_accuracy_score', ascending=False) 236 | if not isinstance(y_test, str): 237 | data_frame = model_comparison_df.set_index('model_name').astype(float) 238 | plt.figure(figsize=(10, 10)) 239 | g = sns.heatmap(data_frame, annot=True, fmt='0.2f', cbar=False) 240 | g.set_xticklabels(g.get_xticklabels(), rotation=45, fontsize=12) 241 | g.set_yticklabels(g.get_yticklabels(), rotation=0, fontsize=12) 242 | print('Time taken = %0.0f seconds' % (time.time() - start_time)) 243 | g.set_title('QuickML Ensembling Models Results', fontsize=18) 244 | f1_stats = dict(estim_tuples) 245 | try: 246 | if scoring in ['logloss', 'rmse', 'mae', 'mape', 'RMSE', 'neg_mean_squared_error']: 247 | best_model_name = min(f1_stats.items(), key=operator.itemgetter(1))[0] 248 | else: 249 | best_model_name = max(f1_stats.items(), key=operator.itemgetter(1))[0] 250 | print('Based on trying multiple models, Best type of algorithm for this data set is %s' % best_model_name) 251 | except: 252 | print('Could not detect best algorithm type from ensembling. Continuing...') 253 | return model_name, stacks 254 | ######################################################### 255 | -------------------------------------------------------------------------------- /autoviml/QuickML_Stacking.py: -------------------------------------------------------------------------------- 1 | import warnings 2 | 3 | import numpy as np 4 | import pandas as pd 5 | 6 | warnings.filterwarnings("ignore") 7 | from sklearn.model_selection import StratifiedKFold, KFold 8 | from sklearn.discriminant_analysis import LinearDiscriminantAnalysis 9 | from sklearn.svm import LinearSVR 10 | from sklearn.tree import DecisionTreeRegressor 11 | from sklearn.linear_model import LassoCV 12 | from sklearn.model_selection import cross_val_predict 13 | from sklearn.naive_bayes import GaussianNB, MultinomialNB 14 | from sklearn.tree import DecisionTreeClassifier 15 | import time 16 | import copy 17 | from collections import Counter 18 | from collections import OrderedDict 19 | 20 | 21 | ############################################################################# 22 | def find_rare_class(classes, verbose=0): 23 | ######### Print the % count of each class in a Target variable ##### 24 | """ 25 | Works on Multi Class too. Prints class percentages count of target variable. 26 | It returns the name of the Rare class (the one with the minimum class member count). 27 | This can also be helpful in using it as pos_label in Binary and Multi Class problems. 28 | """ 29 | counts = OrderedDict(Counter(classes)) 30 | total = sum(counts.values()) 31 | if verbose >= 1: 32 | print(' Class -> Counts -> Percent') 33 | for cls in counts.keys(): 34 | print("%6s: % 7d -> % 5.1f%%" % (cls, counts[cls], counts[cls] / total * 100)) 35 | if type(pd.Series(counts).idxmin()) == str: 36 | return pd.Series(counts).idxmin() 37 | else: 38 | return int(pd.Series(counts).idxmin()) 39 | 40 | 41 | ################################################################################ 42 | def QuickML_Stacking(X_train, y_train, X_test='', modeltype='Regression', Boosting_Flag=False, 43 | scoring='', verbose=0): 44 | """ 45 | Quickly build Stacks of multiple model results 46 | Input must be a clean data set (only numeric variables, no categorical or string variables). 47 | """ 48 | X_train = copy.deepcopy(X_train) 49 | X_test = copy.deepcopy(X_test) 50 | y_train = copy.deepcopy(y_train) 51 | start_time = time.time() 52 | seed = 99 53 | if len(X_train) <= 100000 or X_train.shape[1] < 50: 54 | NUMS = 100 55 | FOLDS = 5 56 | else: 57 | NUMS = 200 58 | FOLDS = 10 59 | ## create Stacking models 60 | estimators = [] 61 | ### This keeps tracks of the number of predict_proba columns generated by each model #### 62 | estimator_length = [] 63 | if isinstance(X_test, str): 64 | no_fit = True 65 | else: 66 | no_fit = False 67 | if no_fit: 68 | #### This is where you don't fit the model but just do cross_val_predict #### 69 | if modeltype == 'Regression': 70 | if scoring == '': 71 | scoring = 'neg_mean_squared_error' 72 | scv = KFold(n_splits=FOLDS, shuffle=False) 73 | if Boosting_Flag: 74 | ###### Bagging models if Bagging is chosen #### 75 | # model4 = BaggingRegressor(DecisionTreeRegressor(random_state=seed), 76 | # n_estimators=NUMS,random_state=seed) 77 | model4 = LinearSVR() 78 | results = cross_val_predict(model4, X_train, y_train, cv=scv, n_jobs=-1) 79 | estimators.append(('Linear_SVR', model4)) 80 | estimator_length.append(1) 81 | elif Boosting_Flag is None: 82 | #### Tree models if Linear chosen ##### 83 | model5 = DecisionTreeRegressor(random_state=seed, min_samples_leaf=2) 84 | results = cross_val_predict(model5, X_train, y_train, cv=scv, n_jobs=-1) 85 | estimators.append(('Decision Trees', model5)) 86 | estimator_length.append(1) 87 | else: 88 | #### Linear Models if Boosting is chosen ##### 89 | model6 = LassoCV(alphas=np.logspace(-5, -1, 20), cv=scv, random_state=seed) 90 | results = cross_val_predict(model6, X_train, y_train, cv=scv, n_jobs=-1) 91 | estimators.append(('LassoCV Regularization', model6)) 92 | estimator_length.append(1) 93 | else: 94 | n_classes = len(Counter(y_train)) 95 | if n_classes > 2: 96 | #### In multi-class setting, it makes sense to turn it into binary class in stage-1 97 | #### In stage 2, a complex model will take the inputs of this model and try to predict 98 | rare_class = find_rare_class(y_train) 99 | if rare_class == 0: 100 | majority_class = 1 101 | else: 102 | majority_class = 0 103 | y_train = y_train.map(lambda x: rare_class if x == rare_class else majority_class) 104 | if scoring == '': 105 | scoring = 'accuracy' 106 | scv = StratifiedKFold(n_splits=FOLDS, random_state=seed, shuffle=True) 107 | if Boosting_Flag: 108 | #### Linear Models if Boosting is chosen ##### 109 | model4 = LinearDiscriminantAnalysis() 110 | results = cross_val_predict(model4, X_train, y_train, cv=scv, n_jobs=-1, 111 | method='predict_proba') 112 | estimators.append(('Linear Discriminant', model4)) 113 | estimator_length.append(results.shape[1]) 114 | elif Boosting_Flag is None: 115 | #### Tree models if Linear chosen ##### 116 | model6 = DecisionTreeClassifier(min_samples_leaf=2) 117 | results = cross_val_predict(model6, X_train, y_train, cv=scv, n_jobs=-1, 118 | method='predict_proba') 119 | estimators.append(('Decision Tree', model6)) 120 | estimator_length.append(results.shape[1]) 121 | else: 122 | ###### Naive Bayes models if Bagging is chosen #### 123 | if n_classes <= 2: 124 | try: 125 | model7 = GaussianNB() 126 | except: 127 | model7 = DecisionTreeClassifier(min_samples_leaf=2) 128 | else: 129 | try: 130 | model7 = MultinomialNB() 131 | except: 132 | model7 = DecisionTreeClassifier(min_samples_leaf=2) 133 | results = cross_val_predict(model7, X_train, y_train, cv=scv, n_jobs=-1, 134 | method='predict_proba') 135 | estimators.append(('Naive Bayes', model7)) 136 | estimator_length.append(results.shape[1]) 137 | else: 138 | #### This is where you fit the model and then predict ######## 139 | if modeltype == 'Regression': 140 | if scoring == '': 141 | scoring = 'neg_mean_squared_error' 142 | scv = KFold(n_splits=FOLDS, shuffle=False) 143 | if Boosting_Flag: 144 | ###### Bagging models if Bagging is chosen #### 145 | # model4 = BaggingRegressor(DecisionTreeRegressor(random_state=seed), 146 | # n_estimators=NUMS,random_state=seed) 147 | model4 = LinearSVR() 148 | results = model4.fit(X_train, y_train).predict(X_test) 149 | estimators.append(('Linear_SVR', model4)) 150 | estimator_length.append(1) 151 | elif Boosting_Flag is None: 152 | #### Tree models if Linear chosen ##### 153 | model5 = DecisionTreeRegressor(random_state=seed, min_samples_leaf=2) 154 | results = model5.fit(X_train, y_train).predict(X_test) 155 | estimators.append(('Decision Trees', model5)) 156 | estimator_length.append(1) 157 | else: 158 | #### Linear Models if Boosting is chosen ##### 159 | model6 = LassoCV(alphas=np.logspace(-5, -1, 20), cv=scv, random_state=seed) 160 | results = model6.fit(X_train, y_train).predict(X_test) 161 | estimators.append(('LassoCV Regularization', model6)) 162 | estimator_length.append(1) 163 | else: 164 | n_classes = len(Counter(y_train)) 165 | if n_classes > 2: 166 | #### In multi-class setting, it makes sense to turn it into binary class in stage-1 167 | #### In stage 2, a complex model will take the inputs of this model and try to predict 168 | rare_class = find_rare_class(y_train) 169 | if rare_class == 0: 170 | majority_class = 1 171 | else: 172 | majority_class = 0 173 | y_train = y_train.map(lambda x: rare_class if x == rare_class else majority_class) 174 | if scoring == '': 175 | scoring = 'accuracy' 176 | scv = StratifiedKFold(n_splits=FOLDS, random_state=seed, shuffle=True) 177 | if Boosting_Flag: 178 | #### Linear Models if Boosting is chosen ##### 179 | model4 = LinearDiscriminantAnalysis() 180 | results = model4.fit(X_train, y_train).predict_proba(X_test) 181 | estimators.append(('Linear Discriminant', model4)) 182 | estimator_length.append(results.shape[1]) 183 | elif Boosting_Flag is None: 184 | #### Tree models if Linear chosen ##### 185 | model6 = DecisionTreeClassifier(min_samples_leaf=2) 186 | results = model6.fit(X_train, y_train).predict_proba(X_test) 187 | estimators.append(('Decision Tree', model6)) 188 | estimator_length.append(results.shape[1]) 189 | else: 190 | ###### Naive Bayes models if Bagging is chosen #### 191 | if n_classes <= 2: 192 | try: 193 | model7 = GaussianNB() 194 | except: 195 | model7 = DecisionTreeClassifier(min_samples_leaf=2) 196 | else: 197 | try: 198 | model7 = MultinomialNB() 199 | except: 200 | model7 = DecisionTreeClassifier(min_samples_leaf=2) 201 | results = model7.fit(X_train, y_train).predict_proba(X_test) 202 | estimators.append(('Naive Bayes', model7)) 203 | estimator_length.append(results.shape[1]) 204 | # stacks = np.c_[results1,results2,results3] 205 | estimators_list = [(tuples[0], tuples[1]) for tuples in estimators] 206 | estimator_names = [tuples[0] for tuples in estimators] 207 | #### Here is where we consolidate the estimator names and their results into one common list ### 208 | ls = [] 209 | for x, y in dict(zip(estimator_names, estimator_length)).items(): 210 | els = [x + '_' + str(eachy) for eachy in range(y)] 211 | ls += els 212 | if verbose == 1: 213 | print(' Time taken for Stacking: %0.1f seconds' % (time.time() - start_time)) 214 | return ls, results 215 | ######################################################### 216 | -------------------------------------------------------------------------------- /autoviml/Transform_KM_Features.py: -------------------------------------------------------------------------------- 1 | from collections import Counter 2 | 3 | import numpy as np 4 | from sklearn.cluster import KMeans 5 | 6 | 7 | class KMeansFeaturizer: 8 | """Transforms numeric data into k-means cluster memberships. 9 | From the Feature Engineering Book by Alice Zheng 10 | https://github.com/alicezheng/feature-engineering-book/blob/master/kmeans_featurizer.py 11 | This transformer runs k-means on the input data and converts each data point 12 | into the id of the closest cluster. If a target variable is present, it is 13 | scaled and included as input to k-means in order to derive clusters that 14 | obey the classification boundary as well as group similar points together. 15 | Parameters 16 | ---------- 17 | k: integer, optional, default 100 18 | The number of clusters to group data into. 19 | target_scale: float, [0, infty], optional, default 5.0 20 | The scaling factor for the target variable. Set this to zero to ignore 21 | the target. For classification problems, larger `target_scale` values 22 | will produce clusters that better respect the class boundary. 23 | random_state : integer or numpy.RandomState, optional 24 | This is passed to k-means as the generator used to initialize the 25 | kmeans centers. If an integer is given, it fixes the seed. Defaults to 26 | the global numpy random number generator. 27 | Attributes 28 | ---------- 29 | cluster_centers_ : array, [k, n_features] 30 | Coordinates of cluster centers. n_features does count the target column. 31 | """ 32 | 33 | def __init__(self, k=100, target_scale=5.0, random_state=None): 34 | self.k = k 35 | self.target_scale = target_scale 36 | self.random_state = random_state 37 | self.km_model_ = None 38 | self.km_model = None 39 | self.cluster_centers_ = None 40 | 41 | def fit(self, X, y=None): 42 | """Runs k-means on the input data and find centroids. 43 | If no target is given (`y` is None) then run vanilla k-means on input 44 | `X`. 45 | If target `y` is given, then include the target (weighted by 46 | `target_scale`) as an extra dimension for k-means clustering. In this 47 | case, run k-means twice, first with the target, then an extra iteration 48 | without. 49 | After fitting, the attribute `cluster_centers_` are set to the k-means 50 | centroids in the input space represented by `X`. 51 | Parameters 52 | ---------- 53 | X : array-like or sparse matrix, shape=(n_data_points, n_features) 54 | y : vector of length n_data_points, optional, default None 55 | If provided, will be weighted with `target_scale` and included in 56 | k-means clustering as hint. 57 | """ 58 | n_features = X.shape[1] 59 | if y is None: 60 | # No target variable, just do plain k-means 61 | km_model = KMeans(n_clusters=self.k, 62 | n_init=20, 63 | random_state=self.random_state) 64 | km_model.fit(X) 65 | 66 | self.km_model_ = km_model 67 | self.cluster_centers_ = km_model.cluster_centers_ 68 | return self 69 | 70 | # There is target information. Apply appropriate scaling and include 71 | # into input data to k-means 72 | data_with_target = np.hstack((X, y[:, np.newaxis] * self.target_scale)) 73 | 74 | # Build a pre-training k-means model on data and target 75 | km_model_pretrain = KMeans(n_clusters=self.k, 76 | n_init=20, 77 | random_state=self.random_state) 78 | km_model_pretrain.fit(data_with_target) 79 | 80 | # Run k-means a second time to get the clusters in the original space 81 | # without target info. Initialize using centroids found in pre-training. 82 | # Go through a single iteration of cluster assignment and centroid 83 | # recomputation. 84 | km_model = KMeans(n_clusters=self.k, 85 | init=km_model_pretrain.cluster_centers_[:, :n_features], 86 | n_init=1, 87 | max_iter=1) 88 | km_model.fit(X) 89 | 90 | self.km_model = km_model 91 | self.cluster_centers_ = km_model.cluster_centers_ 92 | return self 93 | 94 | def transform(self, X): 95 | """Output the closest cluster id for each input data point. 96 | Parameters 97 | ---------- 98 | X : array-like or sparse matrix, shape=(n_data_points, n_features) 99 | Returns 100 | ------- 101 | cluster_ids : array, shape[n_data_points,1] 102 | """ 103 | clusters = self.km_model.predict(X) 104 | return clusters[:, np.newaxis] 105 | 106 | def fit_transform(self, X, y=None): 107 | """Runs fit followed by transform. 108 | """ 109 | self.fit(X, y) 110 | return self.transform(X) 111 | 112 | 113 | import copy 114 | 115 | 116 | def Transform_KM_Features(training_data, training_labels, test_data, km_max=0): 117 | seed = 99 118 | preds = list(training_data) 119 | target = training_labels.name 120 | train_index = training_data.index 121 | test_index = test_data.index 122 | if km_max <= 2: 123 | k_max = 2 124 | else: 125 | k_max = copy.deepcopy(km_max) 126 | ### Calculate the target scale here => the higher the number the better for target accuracy 127 | try: 128 | if training_labels.dtype in [np.float64, np.float32, np.float16]: 129 | target_range = float(abs(training_labels.max() - training_labels.min())) 130 | elif training_labels.dtype in [object, bool]: 131 | target_range = int(len(Counter(training_labels)) + 3) 132 | else: 133 | target_range = int(abs(training_labels.max() - training_labels.min())) 134 | except: 135 | target_range = 5.0 136 | kmf = KMeansFeaturizer(k=k_max, target_scale=target_range, random_state=seed) 137 | kmf_hint = kmf.fit(training_data.values, training_labels.values) 138 | ### Just return it with the cluster column => no need to return the data frame ### 139 | training_cluster_features = kmf_hint.transform(training_data.values) 140 | test_cluster_features = kmf_hint.transform(test_data.values) 141 | npx = np.c_[training_data.values, training_labels.values] 142 | training_with_cluster = np.c_[npx, training_cluster_features] 143 | test_with_cluster = np.c_[test_data, test_cluster_features] 144 | ### We are going to just return the cluster values ###### 145 | train_with_cluster_df = training_with_cluster[:, -1] 146 | test_with_cluster_df = test_with_cluster[:, -1] 147 | # train_with_cluster_df = pd.DataFrame(training_with_cluster,index=train_index, 148 | # columns=preds+[target,'cluster']) 149 | # test_with_cluster_df = pd.DataFrame(test_with_cluster,index=test_index, 150 | # columns=preds+['cluster']) 151 | return train_with_cluster_df, test_with_cluster_df 152 | -------------------------------------------------------------------------------- /autoviml/__init__.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | ################################################################################ 3 | # Auto_ViML - Automatically Build Multiple Interpretable ML Models in Single Line of Code 4 | # Python v3.6+ 5 | # Created by Ram Seshadri 6 | # Licensed under Apache License v2 7 | ################################################################################ 8 | # Version 9 | from .__version__ import __version__, __nlp_version__ 10 | from autoviml.Auto_ViML import Auto_ViML 11 | from autoviml.Auto_NLP import Auto_NLP 12 | if __name__ == "__main__": 13 | module_type = 'Running' 14 | else: 15 | module_type = 'Imported' 16 | viml_version_number = __version__ 17 | nlp_version_number = __nlp_version__ 18 | print("""%s Auto_ViML version: %s. Call using: 19 | m, feats, trainm, testm = Auto_ViML(train, target, test, 20 | sample_submission='', 21 | scoring_parameter='', KMeans_Featurizer=False, 22 | hyper_param='RS',feature_reduction=True, 23 | Boosting_Flag='CatBoost', Binning_Flag=False, 24 | Add_Poly=0, Stacking_Flag=False,Imbalanced_Flag=False, 25 | GPU_flag=False, verbose=1) 26 | """ %(module_type, viml_version_number)) 27 | print() 28 | ########################################################################################### 29 | print("""%s Auto_NLP version: %s.. Call using: 30 | train_nlp, test_nlp, nlp_pipeline, predictions = Auto_NLP( 31 | nlp_column, train, test, target, score_type='balanced_accuracy', 32 | modeltype='Classification',top_num_features=200, verbose=0, 33 | build_model=True)""" %(module_type, nlp_version_number)) 34 | ######################################################################## 35 | -------------------------------------------------------------------------------- /autoviml/__version__.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """Specifies the version of the Auto_ViML package.""" 3 | 4 | __title__ = "Auto_ViML" 5 | __author__ = "Ram Seshadri" 6 | __description__ = "Automatically Build Multiple Interpretable ML Models in Single Line of Code" 7 | __url__ = "https://github.com/AutoViML/Auto_ViML.git" 8 | __version__ = "0.2.02" 9 | __nlp_version__ = "0.1.01" 10 | __license__ = "Apache License 2.0" 11 | __copyright__ = "2020-21 Google" 12 | -------------------------------------------------------------------------------- /autoviml/classify_method.py: -------------------------------------------------------------------------------- 1 | import random 2 | 3 | import numpy as np 4 | import pandas as pd 5 | 6 | np.random.seed(99) 7 | random.seed(42) 8 | ################################################################################ 9 | #### The warnings from Sklearn are so annoying that I have to shut it off ####### 10 | import warnings 11 | 12 | warnings.filterwarnings("ignore") 13 | from sklearn.exceptions import DataConversionWarning 14 | 15 | warnings.filterwarnings(action='ignore', category=DataConversionWarning) 16 | 17 | 18 | def warn(*args, **kwargs): 19 | pass 20 | 21 | 22 | warnings.warn = warn 23 | #################################################################################### 24 | from functools import reduce 25 | 26 | 27 | def left_subtract(l1, l2): 28 | lst = [] 29 | for i in l1: 30 | if i not in l2: 31 | lst.append(i) 32 | return lst 33 | 34 | 35 | ################################################################################# 36 | import copy 37 | 38 | 39 | def EDA_find_remove_columns_with_infinity(df, remove=False): 40 | """ 41 | This function finds all columns in a dataframe that have inifinite values (np.inf or -np.inf) 42 | It returns a list of column names. If the list is empty, it means no columns were found. 43 | If remove flag is set, then it returns a smaller dataframe with inf columns removed. 44 | """ 45 | nums = df.select_dtypes(include='number').columns.tolist() 46 | dfx = df[nums] 47 | sum_rows = np.isinf(dfx).values.sum() 48 | add_cols = list(dfx.columns.to_series()[np.isinf(dfx).any()]) 49 | if sum_rows > 0: 50 | print(' there are %d rows and %d columns with infinity in them...' % (sum_rows, len(add_cols))) 51 | if remove: 52 | ### here you need to use df since the whole dataset is involved ### 53 | nocols = [x for x in df.columns if x not in add_cols] 54 | print(" Shape of dataset before %s and after %s removing columns with infinity" % 55 | (df.shape, (df[nocols].shape,))) 56 | return df[nocols] 57 | else: 58 | ## this will be a list of columns with infinity #### 59 | return add_cols 60 | else: 61 | ## this will be an empty list if there are no columns with infinity 62 | return add_cols 63 | 64 | 65 | #################################################################################### 66 | def classify_columns(df_preds, verbose=0): 67 | """ 68 | This actually does Exploratory data analysis - it means this function performs EDA 69 | ###################################################################################### 70 | Takes a dataframe containing only predictors to be classified into various types. 71 | DO NOT SEND IN A TARGET COLUMN since it will try to include that into various columns. 72 | Returns a data frame containing columns and the class it belongs to such as numeric, 73 | categorical, date or id column, boolean, nlp, discrete_string and cols to delete... 74 | ####### Returns a dictionary with 10 kinds of vars like the following: # continuous_vars,int_vars 75 | # cat_vars,factor_vars, bool_vars,discrete_string_vars,nlp_vars,date_vars,id_vars,cols_delete 76 | """ 77 | train = copy.deepcopy(df_preds) 78 | #### If there are 30 chars are more in a discrete_string_var, it is then considered an NLP variable 79 | max_nlp_char_size = 30 80 | max_cols_to_print = 30 81 | print('#######################################################################################') 82 | print('######################## C L A S S I F Y I N G V A R I A B L E S ####################') 83 | print('#######################################################################################') 84 | print('Classifying variables in data set...') 85 | #### Cat_Limit defines the max number of categories a column can have to be called a categorical colum 86 | cat_limit = 35 87 | float_limit = 15 # Make this limit low so that float variables below this limit become cat vars 88 | 89 | def add(a, b): 90 | return a + b 91 | 92 | sum_all_cols = dict() 93 | orig_cols_total = train.shape[1] 94 | # Types of columns 95 | cols_delete = [] 96 | cols_delete = [col for col in list(train) if (len(train[col].value_counts()) == 1) | ( 97 | train[col].isnull().sum() / len(train) >= 0.90)] 98 | inf_cols = EDA_find_remove_columns_with_infinity(train) 99 | mixed_cols = [x for x in list(train) if len(train[x].dropna().apply(type).value_counts()) > 1] 100 | if len(mixed_cols) > 0: 101 | print(' Removing %s column(s) due to mixed data type detected...' % mixed_cols) 102 | cols_delete += mixed_cols 103 | cols_delete += inf_cols 104 | train = train[left_subtract(list(train), cols_delete)] 105 | var_df = pd.Series(dict(train.dtypes)).reset_index(drop=False).rename( 106 | columns={0: 'type_of_column'}) 107 | sum_all_cols['cols_delete'] = cols_delete 108 | 109 | var_df['bool'] = var_df.apply( 110 | lambda x: 1 if x['type_of_column'] in ['bool', 'object'] and len(train[x['index']].value_counts()) == 2 else 0, 111 | axis=1) 112 | string_bool_vars = list(var_df[(var_df['bool'] == 1)]['index']) 113 | sum_all_cols['string_bool_vars'] = string_bool_vars 114 | var_df['num_bool'] = var_df.apply(lambda x: 1 if x['type_of_column'] in [np.uint8, 115 | np.uint16, np.uint32, np.uint64, 116 | 'int8', 'int16', 'int32', 'int64', 117 | 'float16', 'float32', 'float64'] and len( 118 | train[x['index']].value_counts()) == 2 else 0, axis=1) 119 | num_bool_vars = list(var_df[(var_df['num_bool'] == 1)]['index']) 120 | sum_all_cols['num_bool_vars'] = num_bool_vars 121 | ###### This is where we take all Object vars and split them into diff kinds ### 122 | discrete_or_nlp = var_df.apply(lambda x: 1 if x['type_of_column'] in ['object'] and x[ 123 | 'index'] not in string_bool_vars + cols_delete else 0, axis=1) 124 | ######### This is where we figure out whether a string var is nlp or discrete_string var ### 125 | var_df['nlp_strings'] = 0 126 | var_df['discrete_strings'] = 0 127 | var_df['cat'] = 0 128 | var_df['id_col'] = 0 129 | discrete_or_nlp_vars = var_df.loc[discrete_or_nlp == 1]['index'].values.tolist() 130 | copy_discrete_or_nlp_vars = copy.deepcopy(discrete_or_nlp_vars) 131 | if len(discrete_or_nlp_vars) > 0: 132 | for col in copy_discrete_or_nlp_vars: 133 | #### first fill empty or missing vals since it will blowup ### 134 | ### Remember that fillna only works at the dataframe level! 135 | train[[col]] = train[[col]].fillna(' ') 136 | if train[col].map(lambda x: len(x) if type(x) == str else 0).max( 137 | ) >= 50 and len(train[col].value_counts()) >= int(0.9 * len(train)) and col not in string_bool_vars: 138 | var_df.loc[var_df['index'] == col, 'nlp_strings'] = 1 139 | elif train[col].map(lambda x: len(x) if type(x) == str else 0).mean( 140 | ) >= max_nlp_char_size and train[col].map(lambda x: len(x) if type(x) == str else 0).max( 141 | ) < 50 and len(train[col].value_counts() 142 | ) <= int(0.9 * len(train)) and col not in string_bool_vars: 143 | var_df.loc[var_df['index'] == col, 'discrete_strings'] = 1 144 | elif len(train[col].value_counts()) > cat_limit and len( 145 | train[col].value_counts()) <= int(0.9 * len(train)) and col not in string_bool_vars: 146 | var_df.loc[var_df['index'] == col, 'discrete_strings'] = 1 147 | elif len(train[col].value_counts()) > cat_limit and len(train[col].value_counts() 148 | ) == len(train) and col not in string_bool_vars: 149 | var_df.loc[var_df['index'] == col, 'id_col'] = 1 150 | else: 151 | var_df.loc[var_df['index'] == col, 'cat'] = 1 152 | nlp_vars = list(var_df[(var_df['nlp_strings'] == 1)]['index']) 153 | sum_all_cols['nlp_vars'] = nlp_vars 154 | discrete_string_vars = list(var_df[(var_df['discrete_strings'] == 1)]['index']) 155 | sum_all_cols['discrete_string_vars'] = discrete_string_vars 156 | ###### This happens only if a string column happens to be an ID column ####### 157 | #### DO NOT Add this to ID_VARS yet. It will be done later. Don't change it easily... 158 | #### Category DTYPE vars are very special = they can be left as is and not disturbed in Python. ### 159 | var_df['dcat'] = var_df.apply(lambda x: 1 if str(x['type_of_column']) == 'category' else 0, 160 | axis=1) 161 | factor_vars = list(var_df[(var_df['dcat'] == 1)]['index']) 162 | sum_all_cols['factor_vars'] = factor_vars 163 | ######################################################################## 164 | date_or_id = var_df.apply(lambda x: 1 if x['type_of_column'] in [np.uint8, 165 | np.uint16, np.uint32, np.uint64, 166 | 'int8', 'int16', 167 | 'int32', 'int64'] and x[ 168 | 'index'] not in string_bool_vars + num_bool_vars + discrete_string_vars + nlp_vars else 0, 169 | axis=1) 170 | ######### This is where we figure out whether a numeric col is date or id variable ### 171 | var_df['int'] = 0 172 | var_df['date_time'] = 0 173 | ### if a particular column is date-time type, now set it as a date time variable ## 174 | var_df['date_time'] = var_df.apply(lambda x: 1 if x['type_of_column'] in [' 2050: 182 | var_df.loc[var_df['index'] == col, 'id_col'] = 1 183 | else: 184 | try: 185 | pd.to_datetime(train[col], infer_datetime_format=True) 186 | var_df.loc[var_df['index'] == col, 'date_time'] = 1 187 | except: 188 | var_df.loc[var_df['index'] == col, 'id_col'] = 1 189 | else: 190 | if train[col].min() < 1900 or train[col].max() > 2050: 191 | if col not in num_bool_vars: 192 | var_df.loc[var_df['index'] == col, 'int'] = 1 193 | else: 194 | try: 195 | pd.to_datetime(train[col], infer_datetime_format=True) 196 | var_df.loc[var_df['index'] == col, 'date_time'] = 1 197 | except: 198 | if col not in num_bool_vars: 199 | var_df.loc[var_df['index'] == col, 'int'] = 1 200 | else: 201 | pass 202 | int_vars = list(var_df[(var_df['int'] == 1)]['index']) 203 | date_vars = list(var_df[(var_df['date_time'] == 1)]['index']) 204 | id_vars = list(var_df[(var_df['id_col'] == 1)]['index']) 205 | sum_all_cols['int_vars'] = int_vars 206 | copy_date_vars = copy.deepcopy(date_vars) 207 | for date_var in copy_date_vars: 208 | #### This test is to make sure date vars are actually date vars 209 | try: 210 | pd.to_datetime(train[date_var], infer_datetime_format=True) 211 | except: 212 | ##### if not a date var, then just add it to delete it from processing 213 | cols_delete.append(date_var) 214 | date_vars.remove(date_var) 215 | sum_all_cols['date_vars'] = date_vars 216 | sum_all_cols['id_vars'] = id_vars 217 | sum_all_cols['cols_delete'] = cols_delete 218 | ## This is an EXTREMELY complicated logic for cat vars. Don't change it unless you test it many times! 219 | var_df['numeric'] = 0 220 | float_or_cat = var_df.apply(lambda x: 1 if x['type_of_column'] in ['float16', 221 | 'float32', 'float64'] else 0, 222 | axis=1) 223 | ####### We need to make sure there are no categorical vars in float ####### 224 | if len(var_df.loc[float_or_cat == 1]) > 0: 225 | for col in var_df.loc[float_or_cat == 1]['index'].values.tolist(): 226 | if 2 < len(train[col].value_counts()) <= len( 227 | train) and len(train[col].value_counts() 228 | ) <= float_limit: 229 | var_df.loc[var_df['index'] == col, 'cat'] = 1 230 | else: 231 | if col not in (num_bool_vars + factor_vars): 232 | var_df.loc[var_df['index'] == col, 'numeric'] = 1 233 | cat_vars = list(var_df[(var_df['cat'] == 1)]['index']) 234 | continuous_vars = list(var_df[(var_df['numeric'] == 1)]['index']) 235 | 236 | ######## V E R Y I M P O R T A N T ################################################### 237 | cat_vars_copy = copy.deepcopy(factor_vars) 238 | for cat in cat_vars_copy: 239 | if df_preds[cat].dtype == float: 240 | continuous_vars.append(cat) 241 | factor_vars.remove(cat) 242 | var_df.loc[var_df['index'] == cat, 'dcat'] = 0 243 | var_df.loc[var_df['index'] == cat, 'numeric'] = 1 244 | elif len(df_preds[cat].value_counts()) == df_preds.shape[0]: 245 | id_vars.append(cat) 246 | factor_vars.remove(cat) 247 | var_df.loc[var_df['index'] == cat, 'dcat'] = 0 248 | var_df.loc[var_df['index'] == cat, 'id_col'] = 1 249 | 250 | sum_all_cols['factor_vars'] = factor_vars 251 | ##### There are a couple of extra tests you need to do to remove aberrations in cat_vars ### 252 | cat_vars_copy = copy.deepcopy(cat_vars) 253 | for cat in cat_vars_copy: 254 | if df_preds[cat].dtype == float: 255 | continuous_vars.append(cat) 256 | cat_vars.remove(cat) 257 | var_df.loc[var_df['index'] == cat, 'cat'] = 0 258 | var_df.loc[var_df['index'] == cat, 'numeric'] = 1 259 | elif len(df_preds[cat].value_counts()) == df_preds.shape[0]: 260 | id_vars.append(cat) 261 | cat_vars.remove(cat) 262 | var_df.loc[var_df['index'] == cat, 'cat'] = 0 263 | var_df.loc[var_df['index'] == cat, 'id_col'] = 1 264 | sum_all_cols['cat_vars'] = cat_vars 265 | sum_all_cols['continuous_vars'] = continuous_vars 266 | sum_all_cols['id_vars'] = id_vars 267 | ###### This is where you consolidate the numbers ########### 268 | var_dict_sum = dict(zip(var_df.values[:, 0], var_df.values[:, 2:].sum(1))) 269 | for col, sumval in var_dict_sum.items(): 270 | if sumval == 0: 271 | print('%s of type=%s is not classified' % (col, train[col].dtype)) 272 | elif sumval > 1: 273 | print('%s of type=%s is classified into more then one type' % (col, train[col].dtype)) 274 | else: 275 | pass 276 | ##### If there are more than 1000 unique values, then add it to NLP vars ### 277 | copy_discretes = copy.deepcopy(discrete_string_vars) 278 | for each_discrete in copy_discretes: 279 | if train[each_discrete].nunique() >= 1000: 280 | nlp_vars.append(each_discrete) 281 | discrete_string_vars.remove(each_discrete) 282 | elif 100 < train[each_discrete].nunique() < 1000: 283 | pass 284 | else: 285 | ### If it is less than 100 unique values, then make it categorical var 286 | cat_vars.append(each_discrete) 287 | discrete_string_vars.remove(each_discrete) 288 | sum_all_cols['discrete_string_vars'] = discrete_string_vars 289 | sum_all_cols['cat_vars'] = cat_vars 290 | sum_all_cols['nlp_vars'] = nlp_vars 291 | ############### This is where you print all the types of variables ############## 292 | ####### Returns 8 vars in the following order: continuous_vars,int_vars,cat_vars, 293 | ### string_bool_vars,discrete_string_vars,nlp_vars,date_or_id_vars,cols_delete 294 | if verbose == 1: 295 | print(" Number of Numeric Columns = ", len(continuous_vars)) 296 | print(" Number of Integer-Categorical Columns = ", len(int_vars)) 297 | print(" Number of String-Categorical Columns = ", len(cat_vars)) 298 | print(" Number of Factor-Categorical Columns = ", len(factor_vars)) 299 | print(" Number of String-Boolean Columns = ", len(string_bool_vars)) 300 | print(" Number of Numeric-Boolean Columns = ", len(num_bool_vars)) 301 | print(" Number of Discrete String Columns = ", len(discrete_string_vars)) 302 | print(" Number of NLP String Columns = ", len(nlp_vars)) 303 | print(" Number of Date Time Columns = ", len(date_vars)) 304 | print(" Number of ID Columns = ", len(id_vars)) 305 | print(" Number of Columns to Delete = ", len(cols_delete)) 306 | if verbose == 2: 307 | print(' Printing upto %d columns max in each category:' % max_cols_to_print) 308 | print(" Numeric Columns : %s" % continuous_vars[:max_cols_to_print]) 309 | print(" Integer-Categorical Columns: %s" % int_vars[:max_cols_to_print]) 310 | print(" String-Categorical Columns: %s" % cat_vars[:max_cols_to_print]) 311 | print(" Factor-Categorical Columns: %s" % factor_vars[:max_cols_to_print]) 312 | print(" String-Boolean Columns: %s" % string_bool_vars[:max_cols_to_print]) 313 | print(" Numeric-Boolean Columns: %s" % num_bool_vars[:max_cols_to_print]) 314 | print(" Discrete String Columns: %s" % discrete_string_vars[:max_cols_to_print]) 315 | print(" NLP text Columns: %s" % nlp_vars[:max_cols_to_print]) 316 | print(" Date Time Columns: %s" % date_vars[:max_cols_to_print]) 317 | print(" ID Columns: %s" % id_vars[:max_cols_to_print]) 318 | print(" Columns that will not be considered in modeling: %s" % cols_delete[:max_cols_to_print]) 319 | ##### now collect all the column types and column names into a single dictionary to return! 320 | 321 | len_sum_all_cols = reduce(add, [len(v) for v in sum_all_cols.values()]) 322 | if len_sum_all_cols == orig_cols_total: 323 | print(' %d Predictors classified...' % orig_cols_total) 324 | # print(' This does not include the Target column(s)') 325 | else: 326 | print('No of columns classified %d does not match %d total cols. Continuing...' % ( 327 | len_sum_all_cols, orig_cols_total)) 328 | ls = sum_all_cols.values() 329 | flat_list = [item for sublist in ls for item in sublist] 330 | if len(left_subtract(list(train), flat_list)) == 0: 331 | print(' Missing columns = None') 332 | else: 333 | print(' Missing columns = %s' % left_subtract(list(train), flat_list)) 334 | return sum_all_cols 335 | #################################################################################### 336 | -------------------------------------------------------------------------------- /autoviml/custom_scores.py: -------------------------------------------------------------------------------- 1 | import warnings 2 | 3 | import numpy as np 4 | from sklearn.metrics import accuracy_score, confusion_matrix, f1_score, fbeta_score, log_loss 5 | from sklearn.metrics import mean_squared_error, median_absolute_error, mean_absolute_error 6 | from sklearn.metrics import mean_squared_log_error, precision_score, recall_score, roc_auc_score 7 | 8 | 9 | def balanced_accuracy_score(y_true, y_pred, sample_weight=None, 10 | adjusted=False): 11 | matrix = confusion_matrix(y_true, y_pred, sample_weight=sample_weight) 12 | with np.errstate(divide='ignore', invalid='ignore'): 13 | per_class = np.diag(matrix) / matrix.sum(axis=1) 14 | if np.any(np.isnan(per_class)): 15 | warnings.warn('y_pred contains classes not in y_true') 16 | per_class = per_class[~np.isnan(per_class)] 17 | score = np.mean(per_class) 18 | if adjusted: 19 | n_classes = len(per_class) 20 | chance = 1 / n_classes 21 | score -= chance 22 | score /= 1 - chance 23 | return score 24 | 25 | 26 | ##################################################################################### 27 | def accu(results, y_cv): 28 | return (results == y_cv).astype(int).sum(axis=0) / (y_cv.shape[0]) 29 | 30 | 31 | def rmse(results, y_cv): 32 | return np.sqrt(np.mean((results - y_cv) ** 2, axis=0)) 33 | 34 | 35 | # Defining objective functions for HyperOpt here 36 | def gini(truth, predictions): 37 | g = np.asarray(np.c_[truth, predictions, np.arange(len(truth))], dtype=np.float) 38 | g = g[np.lexsort((g[:, 2], -1 * g[:, 1]))] 39 | gs = g[:, 0].cumsum().sum() / g[:, 0].sum() 40 | gs -= (len(truth) + 1) / 2. 41 | return gs / len(truth) 42 | 43 | 44 | def gini_sklearn(truth, predictions): 45 | return gini(truth, predictions) / gini(truth, truth) 46 | 47 | 48 | def gini_meae(truth, predictions): 49 | score = median_absolute_error(truth, predictions) 50 | return score 51 | 52 | 53 | def gini_msle(truth, predictions): 54 | score = mean_squared_log_error(truth, predictions) 55 | return score 56 | 57 | 58 | def gini_mae(truth, predictions): 59 | score = mean_absolute_error(truth, predictions) 60 | return score 61 | 62 | 63 | def gini_mse(truth, predictions): 64 | score = mean_squared_error(truth, predictions) 65 | return score 66 | 67 | 68 | def gini_rmse(truth, predictions): 69 | score = np.sqrt(mean_squared_error(truth, predictions)) 70 | return score 71 | 72 | 73 | def gini_accuracy(truth, predictions): 74 | return accuracy_score(truth, predictions) 75 | 76 | 77 | def gini_bal_accuracy(truth, predictions): 78 | return balanced_accuracy_score(truth, predictions) 79 | 80 | 81 | def gini_roc(truth, predictions): 82 | return roc_auc_score(truth, predictions) 83 | 84 | 85 | def gini_precision(truth, predictions, pos_label=1): 86 | return precision_score(truth, predictions, average=None)[pos_label] 87 | 88 | 89 | def gini_average_precision(truth, predictions): 90 | return precision_score(truth, predictions.argmax(axis=1), average='macro') 91 | 92 | 93 | def gini_weighted_precision(truth, predictions): 94 | return precision_score(truth, predictions.argmax(axis=1), average='weighted') 95 | 96 | 97 | def gini_macro_precision(truth, predictions): 98 | return precision_score(truth, predictions.argmax(axis=1), average='macro') 99 | 100 | 101 | def gini_micro_precision(truth, predictions): 102 | return precision_score(truth, predictions.argmax(axis=1), average='micro') 103 | 104 | 105 | def gini_samples_precision(truth, predictions): 106 | return precision_score(truth, predictions.argmax(axis=1), average='samples') 107 | 108 | 109 | def gini_f1(truth, predictions, pos_label=1): 110 | return f1_score(truth, predictions, average=None)[pos_label] 111 | 112 | 113 | def gini_weighted_f1(truth, predictions): 114 | return f1_score(truth, predictions.argmax(axis=1), average='weighted') 115 | 116 | 117 | def gini_macro_f1(truth, predictions): 118 | return f1_score(truth, predictions.argmax(axis=1), average='macro') 119 | 120 | 121 | def gini_micro_f1(truth, predictions): 122 | return f1_score(truth, predictions.argmax(axis=1), average='micro') 123 | 124 | 125 | def gini_samples_f1(truth, predictions): 126 | return f1_score(truth, predictions.argmax(axis=1), average='samples') 127 | 128 | 129 | def gini_log_loss(truth, predictions): 130 | return log_loss(truth, predictions, normalize=True) 131 | 132 | 133 | def gini_recall(truth, predictions, pos_label=1): 134 | return recall_score(truth, predictions, average=None)[pos_label] 135 | 136 | 137 | def gini_weighted_recall(truth, predictions): 138 | return recall_score(truth, predictions.argmax(axis=1), average='weighted') 139 | 140 | 141 | def gini_samples_recall(truth, predictions): 142 | return recall_score(truth, predictions.argmax(axis=1), average='samples') 143 | 144 | 145 | def gini_macro_recall(truth, predictions): 146 | return recall_score(truth, predictions.argmax(axis=1), average='macro') 147 | 148 | 149 | def gini_micro_recall(truth, predictions): 150 | return recall_score(truth, predictions.argmax(axis=1), average='micro') 151 | 152 | 153 | # calculate f2-measure 154 | def f2_measure(y_true, y_pred): 155 | return fbeta_score(y_true, y_pred, beta=2) 156 | -------------------------------------------------------------------------------- /autoviml/custom_scores_HO.py: -------------------------------------------------------------------------------- 1 | import warnings 2 | 3 | import numpy as np 4 | from sklearn.metrics import accuracy_score 5 | from sklearn.metrics import average_precision_score 6 | ##################################################################################### 7 | from sklearn.metrics import confusion_matrix 8 | from sklearn.metrics import f1_score 9 | from sklearn.metrics import fbeta_score 10 | from sklearn.metrics import log_loss 11 | from sklearn.metrics import make_scorer 12 | from sklearn.metrics import mean_squared_error, median_absolute_error, mean_absolute_error 13 | from sklearn.metrics import mean_squared_log_error 14 | from sklearn.metrics import precision_score 15 | from sklearn.metrics import recall_score 16 | from sklearn.metrics import roc_auc_score 17 | 18 | 19 | def balanced_accuracy_score(y_true, y_pred, sample_weight=None, 20 | adjusted=False): 21 | C = confusion_matrix(y_true, y_pred, sample_weight=sample_weight) 22 | with np.errstate(divide='ignore', invalid='ignore'): 23 | per_class = np.diag(C) / C.sum(axis=1) 24 | if np.any(np.isnan(per_class)): 25 | warnings.warn('y_pred contains classes not in y_true') 26 | per_class = per_class[~np.isnan(per_class)] 27 | score = np.mean(per_class) 28 | if adjusted: 29 | n_classes = len(per_class) 30 | chance = 1 / n_classes 31 | score -= chance 32 | score /= 1 - chance 33 | return score 34 | 35 | 36 | ##################################################################################### 37 | def accu(results, y_cv): 38 | return (results == y_cv).astype(int).sum(axis=0) / (y_cv.shape[0]) 39 | 40 | 41 | def rmse(results, y_cv): 42 | return np.sqrt(np.mean((results - y_cv) ** 2, axis=0)) 43 | 44 | 45 | ######## Defining objective functions for HyperOpt here ###################### 46 | ### keep all Classification Scorers greater_is_better True but subtract scorer from 1 so it diminishes. 47 | #### This is the only way HyperOpt will find the minimum - so don't change this code anytime! 48 | def gini(truth, predictions): 49 | g = np.asarray(np.c_[truth, predictions, np.arange(len(truth))], dtype=np.float) 50 | g = g[np.lexsort((g[:, 2], -1 * g[:, 1]))] 51 | gs = g[:, 0].cumsum().sum() / g[:, 0].sum() 52 | gs -= (len(truth) + 1) / 2. 53 | return gs / len(truth) 54 | 55 | 56 | def gini_sklearn(truth, predictions): 57 | return gini(truth, predictions) / gini(truth, truth) 58 | 59 | 60 | def gini_meae(truth, predictions): 61 | score = median_absolute_error(truth, predictions) 62 | return score 63 | 64 | 65 | def gini_msle(truth, predictions): 66 | score = np.sqrt(mean_squared_log_error(truth, predictions)) 67 | return score 68 | 69 | 70 | def gini_mae(truth, predictions): 71 | score = mean_absolute_error(truth, predictions) 72 | return score 73 | 74 | 75 | def gini_mse(truth, predictions): 76 | score = np.sqrt(mean_squared_error(truth, predictions)) 77 | return score 78 | 79 | 80 | def gini_rmse(truth, predictions): 81 | score = np.sqrt(mean_squared_error(truth, predictions)) 82 | return score 83 | 84 | 85 | def gini_accuracy(truth, predictions): 86 | return 1 - accuracy_score(truth, predictions) 87 | 88 | 89 | def gini_bal_accuracy(truth, predictions): 90 | return 1 - balanced_accuracy_score(truth, predictions) 91 | 92 | 93 | def gini_roc(truth, predictions): 94 | return 1 - roc_auc_score(truth, predictions) 95 | 96 | 97 | def gini_precision(truth, predictions): 98 | return 1 - precision_score(truth, predictions) 99 | 100 | 101 | def gini_average_precision(truth, predictions): 102 | return 1 - average_precision_score(truth, predictions.argmax(axis=1), average='weighted') 103 | 104 | 105 | def gini_weighted_precision(truth, predictions): 106 | return 1 - precision_score(truth, predictions.argmax(axis=1), average='weighted') 107 | 108 | 109 | def gini_macro_precision(truth, predictions): 110 | return 1 - precision_score(truth, predictions.argmax(axis=1), average='macro') 111 | 112 | 113 | def gini_micro_precision(truth, predictions): 114 | return 1 - precision_score(truth, predictions.argmax(axis=1), average='micro') 115 | 116 | 117 | def gini_samples_precision(truth, predictions): 118 | return 1 - precision_score(truth, predictions.argmax(axis=1), average='samples') 119 | 120 | 121 | def gini_f1(truth, predictions): 122 | return 1 - f1_score(truth, predictions) 123 | 124 | 125 | def gini_weighted_f1(truth, predictions): 126 | return 1 - f1_score(truth, predictions.argmax(axis=1), average='weighted') 127 | 128 | 129 | def gini_macro_f1(truth, predictions): 130 | return 1 - f1_score(truth, predictions.argmax(axis=1), average='macro') 131 | 132 | 133 | def gini_micro_f1(truth, predictions): 134 | return 1 - f1_score(truth, predictions.argmax(axis=1), average='micro') 135 | 136 | 137 | def gini_samples_f1(truth, predictions): 138 | return 1 - f1_score(truth, predictions.argmax(axis=1), average='samples') 139 | 140 | 141 | def gini_log_loss(truth, predictions): 142 | return log_loss(truth, predictions, normalize=True) 143 | 144 | 145 | def gini_recall(truth, predictions): 146 | return 1 - recall_score(truth, predictions) 147 | 148 | 149 | def gini_weighted_recall(truth, predictions): 150 | return 1 - recall_score(truth, predictions.argmax(axis=1), average='weighted') 151 | 152 | 153 | def gini_samples_recall(truth, predictions): 154 | return 1 - recall_score(truth, predictions.argmax(axis=1), average='samples') 155 | 156 | 157 | def gini_macro_recall(truth, predictions): 158 | return 1 - recall_score(truth, predictions.argmax(axis=1), average='macro') 159 | 160 | 161 | def gini_micro_recall(truth, predictions): 162 | return 1 - recall_score(truth, predictions.argmax(axis=1), average='micro') 163 | 164 | 165 | def gini_roc_auc(truth, predictions): 166 | return 1 - roc_auc_score(truth, predictions.argmax(axis=1), average='macro') 167 | 168 | 169 | # calculate f2-measure 170 | def f2_measure(y_true, y_pred): 171 | return fbeta_score(y_true, y_pred, beta=2) 172 | 173 | 174 | ### keep all Regression Scorers greater_is_better True since it leaves them as is and minimizes them 175 | meae_scorer = make_scorer(gini_meae, greater_is_better=True) 176 | msle_scorer = make_scorer(gini_msle, greater_is_better=True) 177 | mae_scorer = make_scorer(gini_mae, greater_is_better=True) 178 | mse_scorer = make_scorer(gini_mse, greater_is_better=True) 179 | 180 | ### keep all Classification Scorers greater_is_better True but subtract scorer from 1 so it diminishes. 181 | #### This is the only way HyperOpt will find the minimum - so don't change this code anytime! 182 | accuracy_scorer = make_scorer(gini_accuracy, greater_is_better=True, needs_proba=False) 183 | bal_accuracy_scorer = make_scorer(gini_bal_accuracy, greater_is_better=True, needs_proba=False) 184 | 185 | gini_scorer = make_scorer(gini_sklearn, greater_is_better=True, needs_proba=True) 186 | roc_scorer = make_scorer(gini_roc, greater_is_better=True, needs_proba=True) 187 | 188 | precision_scorer = make_scorer(gini_precision, greater_is_better=True, needs_proba=False) 189 | average_precision_scorer = make_scorer(gini_average_precision, greater_is_better=True, needs_proba=True) 190 | weighted_precision_scorer = make_scorer(gini_weighted_precision, greater_is_better=True, needs_proba=True) 191 | macro_precision_scorer = make_scorer(gini_macro_precision, greater_is_better=True, needs_proba=True) 192 | micro_precision_scorer = make_scorer(gini_micro_precision, greater_is_better=True, needs_proba=True) 193 | samples_precision_scorer = make_scorer(gini_samples_precision, greater_is_better=True, needs_proba=True) 194 | 195 | f1_scorer = make_scorer(gini_f1, greater_is_better=True, needs_proba=False) 196 | weighted_f1_scorer = make_scorer(gini_weighted_f1, greater_is_better=True, needs_proba=True) 197 | macro_f1_scorer = make_scorer(gini_macro_f1, greater_is_better=True, needs_proba=True) 198 | micro_f1_scorer = make_scorer(gini_micro_f1, greater_is_better=True, needs_proba=True) 199 | samples_f1_scorer = make_scorer(gini_samples_f1, greater_is_better=True, needs_proba=True) 200 | 201 | recall_scorer = make_scorer(gini_recall, greater_is_better=True, needs_proba=False) 202 | weighted_recall_scorer = make_scorer(gini_weighted_recall, greater_is_better=True, needs_proba=True) 203 | macro_recall_scorer = make_scorer(gini_macro_recall, greater_is_better=True, needs_proba=True) 204 | micro_recall_scorer = make_scorer(gini_micro_recall, greater_is_better=True, needs_proba=True) 205 | samples_recall_scorer = make_scorer(gini_samples_recall, greater_is_better=True, needs_proba=True) 206 | 207 | roc_auc_scorer = make_scorer(gini_roc_auc, greater_is_better=True, needs_proba=True) 208 | ### Leave the log-loss scorer as greater_is_better True since it keeps sign and minimizes it. 209 | logloss_scorer = make_scorer(gini_log_loss, greater_is_better=True, needs_proba=True) 210 | ########################################################################################## 211 | -------------------------------------------------------------------------------- /autoviml/feature_engineering.py: -------------------------------------------------------------------------------- 1 | import pandas as pd 2 | import numpy as np 3 | 4 | ######### NEW And FAST WAY to ADD Feature Engg COLUMNS IN A DATA SET ####### 5 | ### Modify Dataframe by adding Features using Feature Tools #### 6 | def add_computational_primitive_features(df, add_types: list, idcolumn=''): 7 | """ 8 | ### Modify Dataframe by adding computational primitive Features using Feature Tools #### 9 | ### What are computational primitives? they are to add, subtract, multiply and divide features 10 | ### Inputs: 11 | ### df: Just sent in the data frame df that you want features added to 12 | ### add_types: list of computational types: 'add', 'subtract', 'multiply' and 'divide'. Choose any or all. 13 | ### idcolumn: this is to create an index for the dataframe since FT runs on index variable. You can leave it empty string. 14 | """ 15 | try: 16 | import featuretools as ft 17 | except: 18 | print('pip install featuretools and try this again') 19 | df = copy.deepcopy(df) 20 | projectid = 'project_prediction' 21 | dataid = 'project_data' 22 | if idcolumn == '': 23 | indexid = 'index' 24 | make_index = True 25 | else: 26 | indexid = idcolumn 27 | make_index = False 28 | # Make an entityset and add the entity 29 | es = ft.EntitySet(id=projectid) 30 | es.entity_from_dataframe(entity_id=dataid, dataframe=df, 31 | make_index=make_index, index=indexid) 32 | 33 | # Run deep feature synthesis with given input primitives or automatically deep 2 34 | if len(add_types) > 0: 35 | ### Build Features based on given primitive types, add_types which is a list 36 | df_mod, feature_defs = ft.dfs(entityset=es, target_entity=dataid, 37 | trans_primitives=add_types) 38 | else: 39 | ### Perform Deep Feature Synthesis Automatically for Depth 2 40 | df_mod, feature_defs = ft.dfs(entityset=es, target_entity=dataid, 41 | max_depth=2, n_jobs=-1, 42 | verbose=0) 43 | if make_index: 44 | df_mod = df_mod.reset_index(drop=True) 45 | return df_mod 46 | 47 | 48 | def feature_engineering(df, ft_requests, idcol): 49 | """ 50 | The Feature Engineering module needs FeatureTools installed to work. 51 | So please do "pip install featuretools" before trying out this module. 52 | It takes a given data set, df and adds features based on the requet types in 53 | ft_requests which can be 'add','subtract','multiply','divide'. If you have 54 | an id_column in the data set, you can provide it as idcol (a string variable). 55 | It will return your modified dataset with 'idcol' as the index. Make sure 56 | you reset the index if you want to return it to its former state. 57 | Also do not send in your entire dataframe! Just send a small dataframe with a few variables. 58 | Once you see how it adds to performance of model, you can add more variables to dataframe. 59 | """ 60 | df = copy.deepcopy(df) 61 | if df.shape[1] < 2: 62 | print('More than one column in dataframe required to perform feature engineering. Returning') 63 | return df 64 | ft_dict = dict(zip(['add', 'multiply', 'subtract', 'divide'], 65 | ['add_numeric', 'multiply_numeric', 66 | 'subtract_numeric', 'divide_numeric'])) 67 | if len(ft_requests) > 0: 68 | ft_list = [] 69 | for ft_one in ft_requests: 70 | if ft_one in ft_dict.keys(): 71 | ft_list.append(ft_dict[ft_one]) 72 | else: 73 | print(' Cannot perform %s-type feature engineering...' % ft_one) 74 | cols = [x for x in df.columns.tolist() if x not in [idcol]] 75 | for each_ft, count in zip(ft_list, range(len(ft_list))): 76 | if count == 0: 77 | df_mod = add_computational_primitive_features(df, [each_ft], idcol) 78 | print(df_mod.shape) 79 | else: 80 | df_temp = add_computational_primitive_features(df, [each_ft], idcol) 81 | df_temp.drop(cols, axis=1, inplace=True) 82 | df_mod = pd.concat([df_mod, df_temp], axis=1, ignore_index=False) 83 | print(df_mod.shape) 84 | else: 85 | df_mod = add_computational_primitive_features(df, [], idcol) 86 | return df_mod 87 | 88 | 89 | def add_date_time_features(smalldf, startTime, endTime, splitter_date_string="/", splitter_hour_string=":"): 90 | """ 91 | If you have start date time stamp and end date time stamp, this module will create additional features for such fields. 92 | You must provide a start date time stamp field and if you have an end date time stamp field, you must use it. 93 | Otherwise, you are better off using the create_date_time_features module which is also in this library. 94 | You must provide the following: 95 | smalldf: Dataframe containing your date time fields 96 | startTime: this is hopefully a string field which converts to a date time stamp easily. Make sure it is a string. 97 | endTime: this also must be a string field which converts to a date time stamp easily. Make sure it is a string. 98 | splitter_date_string: usually there is a string such as '/' or '.' between day/month/year etc. Default is assumed / here. 99 | splitter_hour_string: usually there is a string such as ':' or '.' between hour:min:sec etc. Default is assumed : here. 100 | """ 101 | smalldf = smalldf.copy() 102 | add_cols = [] 103 | start_date = 'processing' + startTime + '_start_date' 104 | smalldf[start_date] = smalldf[startTime].map(lambda x: x.split(" ")[0]) 105 | add_cols.append(start_date) 106 | try: 107 | start_time = 'processing' + startTime + '_start_time' 108 | smalldf[start_time] = smalldf[startTime].map(lambda x: x.split(" ")[1]) 109 | add_cols.append(start_time) 110 | except: 111 | ### there is no hour-minutes part of this date time stamp field. You can just skip it if it is not there 112 | pass 113 | end_date = 'processing' + endTime + '_end_date' 114 | smalldf[end_date] = smalldf[endTime].map(lambda x: x.split(" ")[0]) 115 | add_cols.append(end_date) 116 | try: 117 | end_time = 'processing' + endTime + '_end_time' 118 | smalldf[end_time] = smalldf[endTime].map(lambda x: x.split(" ")[1]) 119 | add_cols.append(end_time) 120 | except: 121 | ### there is no hour-minutes part of this date time stamp field. You can just skip it if it is not there 122 | pass 123 | view_days = 'processing' + startTime + '_elapsed_days' 124 | smalldf[view_days] = (pd.to_datetime(smalldf[end_date]) - pd.to_datetime(smalldf[start_date])).values.astype(int) 125 | add_cols.append(view_days) 126 | try: 127 | view_time = 'processing' + startTime + '_elapsed_time' 128 | smalldf[view_time] = (pd.to_datetime(smalldf[end_time]) - pd.to_datetime(smalldf[start_time])).astype( 129 | 'timedelta64[s]').values 130 | add_cols.append(view_time) 131 | except: 132 | ### In some date time fields this gives an error so skip it in that case 133 | pass 134 | #### The reason we chose endTime here is that startTime is usually taken care of by another library. So better to do this alone. 135 | year = 'processing' + endTime + '_end_year' 136 | smalldf[year] = smalldf[end_date].map(lambda x: str(x).split(splitter_date_string)[0]).values 137 | add_cols.append(year) 138 | #### The reason we chose endTime here is that startTime is usually taken care of by another library. So better to do this alone. 139 | month = 'processing' + endTime + '_end_month' 140 | smalldf[month] = smalldf[end_date].map(lambda x: str(x).split(splitter_date_string)[1]).values 141 | add_cols.append(month) 142 | try: 143 | #### The reason we chose endTime here is that startTime is usually taken care of by another library. So better to do this alone. 144 | daynum = 'processing' + endTime + '_end_day_number' 145 | smalldf[daynum] = smalldf[end_date].map(lambda x: str(x).split(splitter_date_string)[2]).values 146 | add_cols.append(daynum) 147 | except: 148 | ### In some date time fields the day number is not there. If not, just skip it #### 149 | pass 150 | #### In some date time fields, the hour and minute is not there, so skip it in that case if it errors! 151 | try: 152 | start_hour = 'processing' + startTime + '_start_hour' 153 | smalldf[start_hour] = smalldf[start_time].map(lambda x: str(x).split(splitter_hour_string)[0]).values 154 | add_cols.append(start_hour) 155 | start_min = 'processing' + startTime + '_start_hour' 156 | smalldf[start_min] = smalldf[start_time].map(lambda x: str(x).split(splitter_hour_string)[1]).values 157 | add_cols.append(start_min) 158 | except: 159 | ### If it errors, skip it 160 | pass 161 | #### Check if there is a weekday and weekends in date time columns using endTime only 162 | weekday_num = 'processing' + endTime + '_end_weekday_number' 163 | smalldf[weekday_num] = pd.to_datetime(smalldf[end_date]).dt.weekday.values 164 | add_cols.append(weekday_num) 165 | weekend = 'processing' + endTime + '_end_weekend_flag' 166 | smalldf[weekend] = smalldf[weekday_num].map(lambda x: 1 if x in [5, 6] else 0) 167 | add_cols.append(weekend) 168 | #### If everything works well, there should be 13 new columns added by module. All the best! 169 | print('%d columns added using start date=%s and end date=%s processing...' % (len(add_cols), startTime, endTime)) 170 | return smalldf 171 | 172 | ################################################################################# 173 | def split_one_field_into_many(df, field, splitter, filler, new_names_list): 174 | """ 175 | This little function takes any data frame field (string variables only) and splits 176 | it into as many fields as you want in the new_names_list. 177 | You can also specify what string to split on using the splitter argument. 178 | You can also fill Null values that occur due to your splitting by specifying a filler. 179 | if no new_names_list is given, then we use the name of the field itself to split. 180 | """ 181 | import warnings 182 | warnings.filterwarnings("ignore") 183 | df = df.copy() 184 | ### First print the maximum number of things in that field 185 | max_things = df[field].map(lambda x: len(x.split(splitter))).max() 186 | if len(new_names_list) == 0: 187 | print(' Max. columns created by splitting %s field is %d.' % ( 188 | field, max_things)) 189 | else: 190 | print( 191 | ' Max. columns created by splitting %s field is %d but you have given %d variable names only. Selecting first %d' % ( 192 | field, max_things, len(new_names_list), len(new_names_list))) 193 | ### This creates a new field that counts the number of things that are in that field. 194 | num_products_viewed = 'count_things_in_' + field 195 | df[num_products_viewed] = df[field].map(lambda x: len(x.split(";"))).values 196 | ### Clean up the field such that it has the right number of split chars otherwise add to it 197 | df[field] = df[field].map( 198 | lambda x: x + splitter * (max_things - len(x.split(";"))) if len(x.split(";")) < max_things else x) 199 | ###### Now you create new fields by split the one large field ######## 200 | if new_names_list == '': 201 | new_names_list = [field + '_' + str(i) for i in range(1, max_things + 1)] 202 | try: 203 | for i in range(len(new_names_list)): 204 | df[field].fillna(filler, inplace=True) 205 | df.loc[df[field] == splitter, field] = filler 206 | df[new_names_list[i]] = df[field].map(lambda x: x.split(splitter)[i] 207 | if splitter in x else x) 208 | except: 209 | ### Check if the column is a string column. If not, give an error message. 210 | print('Cannot split the column. Getting an error. Check the column again') 211 | return df 212 | return df, new_names_list 213 | 214 | 215 | ################################################################################# 216 | def add_aggregate_primitive_features(dft, agg_types, id_column, ignore_variables: list): 217 | """ 218 | ### Modify Dataframe by adding computational primitive Features using Feature Tools #### 219 | ### What are aggregate primitives? they are to "mean""median","mode","min","max", etc. features 220 | ### Inputs: 221 | ### df: Just sent in the data frame df that you want features added to 222 | ### agg_types: list of computational types: 'mean','median','count', 'max', 'min', 'sum', etc. 223 | ### One caveat: these agg_types must be found in the agg_func of numpy or pandas groupby statement. 224 | ### for example: numpy has 'median','prod','sum','std','var', etc. - they will work! 225 | ### idcolumn: this is to create an index for the dataframe since FT runs on index variable. You can leave it empty string. 226 | ### ignore_variables: list of variables to ignore among numeric variables in data since they may be ID variables. 227 | """ 228 | import copy 229 | ### Make sure the list of functions they send in are acceptable functions. If not, the aggregate will blow up! 230 | func_set = {'count', 'sum', 'mean', 'mad', 'median', 'min', 'max', 'mode', 'abs', 'prod', 'std', 'var', 'sem', 231 | 'skew', 'kurt', 'quantile', 'cumsum', 'cumprod', 'cummax', 'cummin'} 232 | agg_types = list(set(agg_types).intersection(func_set)) 233 | ### If the ignore_variables list is empty, make sure you add the id_column to it so it can be dropped from aggregation. 234 | if len(ignore_variables) == 0: 235 | ignore_variables = [id_column] 236 | ### Select only integer and float variables to do this aggregation on. Be very careful if there are too many vars. 237 | ### This will take time to run in that case. 238 | dft_index = copy.deepcopy(dft[id_column]) 239 | dft_cont = copy.deepcopy(dft.select_dtypes('number').drop(ignore_variables, axis=1)) 240 | dft_cont[id_column] = dft_index 241 | try: 242 | dft_full = dft_cont.groupby(id_column).agg(agg_types) 243 | except: 244 | ### if for some reason, the groupby blows up, then just return the dataframe as is - no changes! 245 | return dft 246 | cols = [x + '_' + y + '_by_' + id_column for (x, y) in dft_full.columns] 247 | dft_full.columns = cols 248 | ### Not every column has useful values. If it is full of just the same value, remove it 249 | _, list_unique_col_ids = np.unique(dft_full, axis=1, return_index=True) 250 | dft_full = dft_full.iloc[:, list_unique_col_ids] 251 | return dft_full 252 | 253 | 254 | ################################################################################################################################ 255 | import copy 256 | import time 257 | import pdb 258 | 259 | 260 | def FE_create_groupby_features(dft, groupby_columns, numeric_columns, agg_types): 261 | """ 262 | FE means FEATURE ENGINEERING - That means this function will create new features 263 | Beware: this function will return a smaller dataframe than what you send in since it groups rows by keys. 264 | ######################################################################################################### 265 | Function groups rows in a dft dataframe by the groupby_columns and returns multiple columns for the numeric column aggregated. 266 | Do not send in more than one column in the numeric column since beyond the first column it will be ignored! 267 | agg_type can be any numpy function such as mean, median, sum, count, etc. 268 | ########################################################################################################## 269 | Returns: a smaller dataframe with rows grouped by groupby_columns and aggregated for the numeric_column 270 | """ 271 | start_time = time.time() 272 | print('Autoviml Feature Engineering: creating groupby features using %s' % groupby_columns) 273 | ########## This is where we create new columns by each numeric column grouped by group-by columns given. 274 | if isinstance(numeric_columns, list): 275 | pass 276 | elif isinstance(numeric_columns, str): 277 | numeric_columns = [numeric_columns] 278 | else: 279 | print(' Numeric column must be a string not a number Try again') 280 | return pd.DataFrame() 281 | grouped_list = pd.DataFrame() 282 | for iteration, numeric_column in zip(range(len(numeric_columns)), numeric_columns): 283 | grouped = dft.groupby(groupby_columns)[[numeric_column]] 284 | try: 285 | agg_type = agg_types[iteration] 286 | except: 287 | print(' No aggregation type given, hence mean is chosen by default') 288 | agg_type = 'mean' 289 | try: 290 | prefix = numeric_column + '_' 291 | if agg_type in ['Sum', 'sum']: 292 | grouped_agg = grouped.sum() 293 | elif agg_type in ['Mean', 'mean', 'Average', 'average']: 294 | grouped_agg = grouped.mean() 295 | elif agg_type in ['count', 'Count']: 296 | grouped_agg = grouped.count() 297 | elif agg_type in ['Median', 'median']: 298 | grouped_agg = grouped.median() 299 | elif agg_type in ['Maximum', 'maximum', 'max', 'Max']: 300 | ## maximum of the amounts 301 | grouped_agg = grouped.max() 302 | elif agg_type in ['Minimum', 'minimum', 'min', 'Min']: 303 | ## maximum of the amounts 304 | grouped_agg = grouped.min() 305 | else: 306 | grouped_agg = grouped.mean() 307 | grouped_sep = grouped_agg.unstack().add_prefix(prefix).fillna(0) 308 | except: 309 | print(' Error in creating groupby features...returning with null dataframe') 310 | grouped_sep = pd.DataFrame() 311 | if iteration == 0: 312 | grouped_list = copy.deepcopy(grouped_sep) 313 | else: 314 | grouped_list = pd.concat([grouped_list, grouped_sep], axis=1) 315 | print( 316 | ' After grouped features added by %s, number of columns = %d' % (numeric_column, grouped_list.shape[1])) 317 | #### once everything is done, you can close it here 318 | print('Time taken for creation of groupby features (in seconds) = %0.0f' % (time.time() - start_time)) 319 | try: 320 | grouped_list.columns = grouped_list.columns.get_level_values(1) 321 | grouped_list.columns.name = None ## make sure the name on columns is removed 322 | grouped_list = grouped_list.reset_index() ## make sure the ID column comes back 323 | except: 324 | print(' Error in setting column names. Please reset column names after this step...') 325 | return grouped_list 326 | ################################################################################ 327 | from scipy.stats import skew, kurtosis 328 | from sklearn.base import BaseEstimator, TransformerMixin 329 | from sklearn.tree import DecisionTreeRegressor, DecisionTreeClassifier 330 | from sklearn.metrics import mean_squared_error, accuracy_score 331 | 332 | class EntropyBinningTransformer(BaseEstimator, TransformerMixin): 333 | """ 334 | ######### This innovative binning transformer is found only in AutoViML ############################## 335 | The entropy binning approach in AutoViML is a novel and intriguing approach, especially as 336 | it leverages decision trees to determine optimal binning thresholds based on information gain 337 | (for classification) or mean squared error reduction (for regression). 338 | 339 | This method can adapt well to non-linear relationships between the feature and the target. 340 | It also incorporates some heuristics and improvements to enhance its effectiveness and efficiency. 341 | 342 | Heuristics: 343 | Variable Selection: We prioritize binning the most important variables. 344 | This is a sensible heuristic as it focuses computational resources on variables that are 345 | most likely to impact the model. We introduce a parameter to specify the number of top variables to bin. 346 | 347 | Adaptive max_depth: We've used an adaptive max_depth based on the number of continuous variables. 348 | This heuristic helps control the granularity of the binning. We can maintain this by adding a 349 | logic to adjust max_depth based on the number of features. 350 | 351 | Handling of Binned Variables: The code has an option to either keep the original variables 352 | alongside the binned versions or replace them. This flexibility can be useful and we keep it 353 | in the class implementation. 354 | 355 | Post-Binning Smoothing: After binning, applying smoothing techniques (e.g., Laplace smoothing) 356 | can help address bins with very low frequencies, which might be particularly useful for categorical targets. 357 | 358 | Incorporating these aspects, this is an updated version of the original EntropyBinningTransformer method in AutoViML 359 | 360 | Additional Improvements to consider: 361 | 362 | Evaluation and Optimization: Integrating a mechanism to evaluate the effectiveness of binning (e.g., 363 | using cross-validation scores) and automatically optimize the binning parameters accordingly. 364 | ######################################################################################################### 365 | EntropyBinningTransformer is a transformer class for binning numeric variables based on entropy using decision trees, 366 | with options for selecting the most relevant variables for binning based on their predictive power, skewness, 367 | and kurtosis, and for smoothing the bins post-binning to mitigate issues with overfitting and bins with 368 | low frequencies. 369 | 370 | Parameters: 371 | - max_depth: The maximum depth of the decision trees used for binning. 372 | - min_samples_leaf: The minimum number of samples required to be at a leaf node of the decision trees. 373 | - entropy_binning: Whether to apply entropy binning. 374 | - modeltype: Indicates whether the target variable is for 'Regression' or 'Classification'. 375 | 376 | Methods: 377 | - fit: Learns the binning thresholds for each variable. 378 | - transform: Applies the learned thresholds to bin the data, with an option for post-binning smoothing. 379 | - select_top_n_vars: Selects the top N variables for binning based on their predictive power and distribution 380 | characteristics. 381 | ######################################################################################################### 382 | #### U S A G E ##### 383 | # Initialize the EntropyBinningTransformer 384 | entropy_binner = EntropyBinningTransformer( replace_vars=True, modeltype='Classification') 385 | 386 | # Optionally, select top n variables based on their predictive power 387 | # This step is useful if you want to bin only the most informative variables 388 | top_vars = entropy_binner.select_top_n_vars(X_train, y_train, n=2) 389 | X_train_top_vars = X_train[top_vars] 390 | 391 | # Fit the transformer to the training data 392 | entropy_binner.fit(X_train_top_vars, y_train) 393 | 394 | # Transform the training data and new data using the learned binning thresholds 395 | X_train_binned = entropy_binner.transform(X_train_top_vars) 396 | X_test_binned = entropy_binner.transform(X_test[top_vars]) 397 | ######################################################################################################### 398 | """ 399 | def __init__(self, replace_vars=True, modeltype='Classification', top_n_vars=None): 400 | self.replace_vars = replace_vars 401 | self.modeltype = modeltype 402 | self.max_depth = 10 403 | self.min_samples_leaf = 2 404 | self.top_n_vars = top_n_vars 405 | self.binning_thresholds = {} 406 | self.remvars = [] 407 | 408 | def fit(self, X, y): 409 | # Determine top_n variables based on importance, for example, using feature importances from a preliminary model 410 | if self.top_n_vars is not None: 411 | top_vars = self.select_top_n_vars(X, y, n=self.top_n_vars) 412 | else: 413 | top_vars = X.columns 414 | 415 | for col in top_vars: 416 | max_depth = self._adjust_max_depth(len(top_vars)) 417 | clf = self.get_decision_tree(max_depth) 418 | 419 | try: 420 | clf.fit(X[col].values.reshape(-1, 1), y) 421 | thresholds = clf.tree_.threshold[clf.tree_.threshold > -2] 422 | self.binning_thresholds[col] = np.sort(thresholds) 423 | except Exception as e: 424 | print(f'Error in {col} during Entropy Binning: {e}') 425 | 426 | return self 427 | 428 | def get_decision_tree(self, max_depth): 429 | if self.modeltype == 'Regression': 430 | return DecisionTreeRegressor(criterion='mse', min_samples_leaf=self.min_samples_leaf, 431 | max_depth=max_depth, random_state=99) 432 | else: 433 | return DecisionTreeClassifier(criterion='entropy', min_samples_leaf=self.min_samples_leaf, 434 | max_depth=max_depth, random_state=99) 435 | 436 | def _adjust_max_depth(self, n_features): 437 | """ 438 | Adjust max depth based on the heuristic from the original function. 439 | 440 | Future: 441 | Dynamic Depth Adjustment: Instead of fixed depth adjustments based on the number of variables, 442 | use a more dynamic approach that considers the actual distribution and variability of each variable. 443 | 444 | """ 445 | if self.max_depth is not None: 446 | return self.max_depth # Use user-defined max_depth if provided 447 | elif n_features <= 2: 448 | return 2 449 | elif n_features <= 5: 450 | return n_features - 2 451 | elif n_features <= 10: 452 | return 5 453 | else: 454 | return 10 # Default max depth for > 10 features 455 | 456 | def select_top_n_vars(self, X, y, n=2, skew_threshold=1.5, kurtosis_threshold=3): 457 | """ 458 | The select_top_n_vars method takes into account the skewness and kurtosis of each variable 459 | in addition to its predictive power. The final score is used to rank the variables 460 | adjusted by the absolute values of skewness and kurtosis to prefer variables that are 461 | both predictive and have less extreme distributions. 462 | 463 | Incorporating distribution-based metrics such as skewness, kurtosis, and outlier 464 | detection into the selection process for variables to bin can indeed provide a more 465 | nuanced approach. Variables with high skewness or kurtosis may benefit more from binning, 466 | as it can help to normalize their effect in predictive models. Additionally, handling 467 | outliers before binning can ensure that bins are more representative of the general 468 | distribution of the data. 469 | """ 470 | scores = {} 471 | ### Select only float numeric columns to bin ########### 472 | numvars = X.select_dtypes(include='float').columns.tolist() 473 | ### Let's keep the remaining vars ###### 474 | self.remvars = [x for x in list(X) if x not in numvars ] 475 | for col in numvars: 476 | X_col = X[col].values.reshape(-1, 1) 477 | 478 | # Calculate skewness and kurtosis 479 | skewness = skew(X[col]) 480 | kurt = kurtosis(X[col], fisher=True) # Fisher's definition is used (normal ==> 0.0) 481 | 482 | # Initialize a simple decision tree 483 | if self.modeltype == 'Regression': 484 | model = DecisionTreeRegressor(max_depth=self.max_depth, random_state=99) 485 | else: 486 | model = DecisionTreeClassifier(max_depth=self.max_depth, random_state=99) 487 | 488 | # Fit the model 489 | model.fit(X_col, y) 490 | 491 | # Predict using the model 492 | predictions = model.predict(X_col) 493 | 494 | # Calculate the performance metric 495 | if self.modeltype == 'Regression': 496 | score = -mean_squared_error(y, predictions) # Negative because lower MSE is better 497 | else: 498 | score = accuracy_score(y, predictions) 499 | 500 | # Store the score, possibly adjusted for skewness and kurtosis 501 | if abs(skewness) > skew_threshold or abs(kurt) > kurtosis_threshold: 502 | scores[col] = score * (1 + abs(skewness) + abs(kurt)) 503 | else: 504 | scores[col] = score 505 | 506 | # Sort the variables based on the scores and select the top n 507 | top_n_vars = sorted(scores, key=scores.get, reverse=True)[:max(n, int(X.shape[1] * 0.1), 2)] 508 | 509 | self.remvars += [x for x in numvars if x not in top_n_vars ] 510 | return top_n_vars 511 | 512 | def transform(self, X): 513 | """ 514 | Applies the learned binning thresholds and Laplace smoothing to the data, replacing or appending the original 515 | variables with their smoothed, binned versions based on the replace_vars flag. 516 | 517 | Parameters: 518 | - X (pd.DataFrame): The input data to transform. 519 | 520 | Returns: 521 | - X_transformed (pd.DataFrame): The transformed data with variables binned and smoothed according to the learned thresholds. 522 | """ 523 | X_transformed = X.copy() 524 | for col, thresholds in self.binning_thresholds.items(): 525 | bin_col_name = f'{col}_bin' if not self.replace_vars else col 526 | binned_data = np.digitize(X_transformed[col].values, thresholds) 527 | 528 | # Apply Laplace smoothing to the binned data 529 | smoothed_binned_data = binned_data + 1 # Add-one smoothing 530 | 531 | # Replace or append the smoothed, binned data 532 | X_transformed[bin_col_name] = smoothed_binned_data 533 | 534 | # If replacing, drop the original column if it's not the same as the bin_col_name 535 | if self.replace_vars and bin_col_name != col: 536 | X_transformed.drop(col, axis=1, inplace=True) 537 | 538 | return X_transformed 539 | ################################################################################################ 540 | -------------------------------------------------------------------------------- /autoviml/sulov_method.py: -------------------------------------------------------------------------------- 1 | import random 2 | 3 | import numpy as np 4 | import pandas as pd 5 | import networkx as nx 6 | 7 | np.random.seed(99) 8 | random.seed(42) 9 | ################################################################################ 10 | #### The warnings from Sklearn are so annoying that I have to shut it off ####### 11 | import warnings 12 | warnings.filterwarnings("ignore") 13 | from sklearn.exceptions import DataConversionWarning 14 | warnings.filterwarnings(action='ignore', category=DataConversionWarning) 15 | def warn(*args, **kwargs): 16 | pass 17 | warnings.warn = warn 18 | #################################################################################### 19 | import copy 20 | from sklearn.feature_selection import mutual_info_regression, mutual_info_classif 21 | from sklearn.feature_selection import SelectKBest 22 | from itertools import combinations 23 | import matplotlib.patches as mpatches 24 | import matplotlib.pyplot as plt 25 | from collections import defaultdict 26 | ################################################################################################# 27 | def remove_highly_correlated_vars_fast(df, corr_limit=0.70): 28 | """ 29 | This is a simple method to remove highly correlated features fast using Pearson's Correlation. 30 | Use this only for float and integer variables. It will automatically select those only. 31 | It can be used for very large data sets where featurewiz has trouble with memory 32 | """ 33 | # Creating correlation matrix 34 | cor_matrix = df.corr().abs().astype(np.float16) 35 | # Selecting upper triangle of correlation matrix 36 | upper_tri = cor_matrix.where(np.triu(np.ones(cor_matrix.shape), 37 | k=1).astype(bool)) 38 | # Finding index of feature columns with correlation greater than 0.95 39 | to_drop = [column for column in upper_tri.columns if any(upper_tri[column] > corr_limit)] 40 | print() 41 | print('Highly correlated columns to remove: %s' % to_drop) 42 | return to_drop 43 | 44 | 45 | ##################################################################################### 46 | def left_subtract(l1, l2): 47 | lst = [] 48 | for i in l1: 49 | if i not in l2: 50 | lst.append(i) 51 | return lst 52 | 53 | ################################################################################# 54 | def return_dictionary_list(lst_of_tuples): 55 | """ Returns a dictionary of lists if you send in a list of Tuples""" 56 | orDict = defaultdict(list) 57 | # iterating over list of tuples 58 | for key, val in lst_of_tuples: 59 | orDict[key].append(val) 60 | return orDict 61 | 62 | 63 | ################################################################################ 64 | def find_remove_duplicates(list_of_values): 65 | """ 66 | # Removes duplicates from a list to return unique values - USED ONLY ONCE 67 | """ 68 | output = [] 69 | seen = set() 70 | for value in list_of_values: 71 | if value not in seen: 72 | output.append(value) 73 | seen.add(value) 74 | return output 75 | 76 | 77 | ################################################################################## 78 | def FE_remove_variables_using_SULOV_method(df, preds_in, modeltype, target, 79 | corr_limit=0.70, verbose=0, dask_xgboost_flag=False): 80 | """ 81 | FE stands for Feature Engineering - it means this function performs feature engineering 82 | ########################################################################################### 83 | ##### SULOV stands for Searching Uncorrelated List Of Variables ############# 84 | This highly efficient method removes variables that are highly correlated using a series of 85 | pair-wise correlation knockout rounds. It is extremely fast and hence can work on thousands 86 | of variables in less than a minute, even on a laptop. You need to send in a list of numeric 87 | variables and that's all! The method defines high Correlation as anything over 0.70 (absolute) 88 | but this can be changed. If two variables have absolute correlation higher than this, they 89 | will be marked, and using a process of elimination, one of them will get knocked out: 90 | To decide order of variables to keep, we use mutuail information score to select. MIS returns 91 | a ranked list of these correlated variables: when we select one, we knock out others 92 | that it is correlated to. Then we select next var. This way we knock out correlated variables. 93 | Finally we are left with uncorrelated variables that are also highly important in mutual score. 94 | ######## YOU MUST INCLUDE THE ABOVE MESSAGE IF YOU COPY THIS CODE IN YOUR LIBRARY ########## 95 | """ 96 | 97 | df = copy.deepcopy(df) 98 | df_target = df[target] 99 | numvars = df[preds_in].select_dtypes(include='float').columns.tolist() 100 | remvars = left_subtract(preds_in, numvars) 101 | df = df[numvars] 102 | ### for some reason, doing a mass fillna of vars doesn't work! Hence doing it individually! 103 | null_vars = np.array(numvars)[df.isnull().sum() > 0] 104 | for each_num in null_vars: 105 | ### Remember that fillna only works at the dataframe level! 106 | df[[each_num]] = df[[each_num]].fillna(0) 107 | target = copy.deepcopy(target) 108 | 109 | print('#######################################################################################') 110 | print('##### SULOV: Searching for Uncorrelated List Of Variables in %s features ############' % len(numvars)) 111 | print('#######################################################################################') 112 | ### This is a shorter version of getting unduplicated and highly correlated vars ## 113 | correlation_dataframe = df.corr().abs().unstack().sort_values().drop_duplicates() 114 | corrdf = pd.DataFrame(correlation_dataframe[:].reset_index()) 115 | corrdf.columns = ['var1', 'var2', 'coeff'] 116 | corrdf1 = corrdf[corrdf['coeff'] >= corr_limit] 117 | correlated_pair = list(zip(corrdf1['var1'].values.tolist(), corrdf1['var2'].values.tolist())) 118 | corr_pair_dict = dict(return_dictionary_list(correlated_pair)) 119 | corr_list = find_remove_duplicates(corrdf1['var1'].values.tolist() + corrdf1['var2'].values.tolist()) 120 | keys_in_dict = list(corr_pair_dict.keys()) 121 | reverse_correlated_pair = [(y, x) for (x, y) in correlated_pair] 122 | reverse_corr_pair_dict = dict(return_dictionary_list(reverse_correlated_pair)) 123 | #### corr_pair_dict is used later to make the network diagram to see which vars are correlated to which 124 | for key, val in reverse_corr_pair_dict.items(): 125 | if key in keys_in_dict: 126 | if len(key) > 1: 127 | corr_pair_dict[key] += val 128 | else: 129 | corr_pair_dict[key] = val 130 | 131 | ###### This is for ordering the variables in the highest to lowest importance to target ### 132 | if len(corr_list) == 0: 133 | final_list = list(correlation_dataframe) 134 | print('Selecting all (%d) variables since none of numeric vars are highly correlated...' % len(numvars)) 135 | return numvars 136 | else: 137 | if isinstance(target, list): 138 | target = target[0] 139 | max_feats = len(corr_list) 140 | if modeltype == 'Regression': 141 | sel_function = mutual_info_regression 142 | fs = SelectKBest(score_func=sel_function, k=max_feats) 143 | else: 144 | sel_function = mutual_info_classif 145 | fs = SelectKBest(score_func=sel_function, k=max_feats) 146 | ##### you must ensure there are no infinite nor null values in corr_list df ## 147 | df_fit = df[corr_list] 148 | ### Now check if there are any NaN values in the dataset ##### 149 | 150 | if df_fit.isnull().sum().sum() > 0: 151 | df_fit = df_fit.dropna() 152 | else: 153 | print(' there are no null values in dataset...') 154 | ##### Reduce memory usage and find mutual information score #### 155 | # try: 156 | # df_fit = reduce_mem_usage(df_fit) 157 | # except: 158 | # print('Reduce memory erroring. Continuing...') 159 | ##### Ready to perform fit and find mutual information score #### 160 | try: 161 | fs.fit(df_fit, df_target) 162 | except: 163 | print(' SelectKBest() function is erroring. Returning with all %s variables...' % len(numvars)) 164 | return numvars 165 | try: 166 | ################################################################################# 167 | ####### This is the main section where we use mutual info score to select vars 168 | ################################################################################# 169 | mutual_info = dict(zip(corr_list, fs.scores_)) 170 | #### The first variable in list has the highest correlation to the target variable ### 171 | sorted_by_mutual_info = [key for (key, val) in 172 | sorted(mutual_info.items(), key=lambda kv: kv[1], reverse=True)] 173 | ##### Now we select the final list of correlated variables ########### 174 | selected_corr_list = [] 175 | #### You have to make multiple copies of this sorted list since it is iterated many times #### 176 | orig_sorted = copy.deepcopy(sorted_by_mutual_info) 177 | copy_sorted = copy.deepcopy(sorted_by_mutual_info) 178 | copy_pair = copy.deepcopy(corr_pair_dict) 179 | #### select each variable by the highest mutual info and see what vars are correlated to it 180 | for each_corr_name in copy_sorted: 181 | ### add the selected var to the selected_corr_list 182 | selected_corr_list.append(each_corr_name) 183 | for each_remove in copy_pair[each_corr_name]: 184 | #### Now remove each variable that is highly correlated to the selected variable 185 | if each_remove in copy_sorted: 186 | copy_sorted.remove(each_remove) 187 | ##### Now we combine the uncorrelated list to the selected correlated list above 188 | rem_col_list = left_subtract(numvars, corr_list) 189 | final_list = rem_col_list + selected_corr_list 190 | removed_cols = left_subtract(numvars, final_list) 191 | except Exception as e: 192 | print(' SULOV Method crashing due to %s' % e) 193 | #### Dropping highly correlated Features fast using simple linear correlation ### 194 | removed_cols = remove_highly_correlated_vars_fast(df, corr_limit) 195 | final_list = left_subtract(numvars, removed_cols) 196 | if len(removed_cols) > 0: 197 | print(' Removing (%d) highly correlated variables:' % (len(removed_cols))) 198 | if len(removed_cols) <= 30: 199 | print(' %s' % removed_cols) 200 | if len(final_list) <= 30: 201 | print(' Following (%d) vars selected: %s' % (len(final_list), final_list)) 202 | ############## D R A W C O R R E L A T I O N N E T W O R K ################## 203 | selected = copy.deepcopy(final_list) 204 | try: 205 | #### Now start building the graph ################### 206 | gf = nx.Graph() 207 | ### the mutual info score gives the size of the bubble ### 208 | multiplier = 2100 209 | for each in orig_sorted: 210 | gf.add_node(each, size=int(max(1, mutual_info[each] * multiplier))) 211 | ######### This is where you calculate the size of each node to draw 212 | sizes = [mutual_info[x] * multiplier for x in list(gf.nodes())] 213 | #### The sizes of the bubbles for each node is determined by its mutual information score value 214 | corr = df_fit.corr() 215 | high_corr = corr[abs(corr) > corr_limit] 216 | ## high_corr is the dataframe of a few variables that are highly correlated to each other 217 | combos = combinations(corr_list, 2) 218 | ### this gives the strength of correlation between 2 nodes ## 219 | multiplier = 20 220 | for (var1, var2) in combos: 221 | if np.isnan(high_corr.loc[var1, var2]): 222 | pass 223 | else: 224 | gf.add_edge(var1, var2, weight=multiplier * high_corr.loc[var1, var2]) 225 | ######## Now start building the networkx graph ########################## 226 | widths = nx.get_edge_attributes(gf, 'weight') 227 | nodelist = gf.nodes() 228 | cols = 5 229 | height_size = 5 230 | width_size = 15 231 | rows = int(len(corr_list) / cols) 232 | if rows < 1: 233 | rows = 1 234 | plt.figure(figsize=(width_size, min(20, height_size * rows))) 235 | pos = nx.shell_layout(gf) 236 | nx.draw_networkx_nodes(gf, pos, 237 | nodelist=nodelist, 238 | node_size=sizes, 239 | node_color='blue', 240 | alpha=0.5) 241 | nx.draw_networkx_edges(gf, pos, 242 | edgelist=widths.keys(), 243 | width=list(widths.values()), 244 | edge_color='lightblue', 245 | alpha=0.6) 246 | pos_higher = {} 247 | x_off = 0.04 # offset on the x axis 248 | y_off = 0.04 # offset on the y axis 249 | for k, v in pos.items(): 250 | pos_higher[k] = (v[0] + x_off, v[1] + y_off) 251 | if len(selected) == 0: 252 | nx.draw_networkx_labels(gf, pos=pos_higher, 253 | labels=dict(zip(nodelist, nodelist)), 254 | font_color='black') 255 | else: 256 | nx.draw_networkx_labels(gf, pos=pos_higher, 257 | labels=dict(zip(nodelist, 258 | [x + ' (selected)' if x in selected else x + ' (removed)' for x 259 | in nodelist])), 260 | font_color='black') 261 | plt.box(True) 262 | plt.title("""In SULOV, we repeatedly remove features with lower mutual info scores among highly correlated pairs (see figure), 263 | SULOV selects the feature with higher mutual info score related to target when choosing between a pair. """, 264 | fontsize=10) 265 | plt.suptitle('How SULOV Method Works by Removing Highly Correlated Features', fontsize=20, y=1.03) 266 | red_patch = mpatches.Patch(color='blue', label='Bigger circle denotes higher mutual info score with target') 267 | blue_patch = mpatches.Patch(color='lightblue', 268 | label='Thicker line denotes higher correlation between two variables') 269 | plt.legend(handles=[red_patch, blue_patch], loc='best') 270 | plt.show() 271 | ##### N E T W O R K D I A G R A M C O M P L E T E ################# 272 | return final_list + remvars 273 | except Exception as e: 274 | print(' Networkx library visualization crashing due to %s' % e) 275 | print('Continuing with SULOV. %d features selected' % len(final_list)) 276 | return final_list + remvars 277 | ################################################################################### 278 | -------------------------------------------------------------------------------- /logo.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AutoViML/Auto_ViML/0761412417b6dac0683425c372fd7e4f8942d815/logo.png -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | numpy<2 2 | pandas>=2.0 3 | ipython 4 | scipy==1.11.4 5 | matplotlib>3.7.4 6 | beautifulsoup4 7 | emoji 8 | ipython 9 | jupyter 10 | seaborn 11 | catboost>=1.2.7 12 | textblob 13 | nltk 14 | regex 15 | xgboost>=0.82,<=1.7.6 16 | vaderSentiment 17 | imbalanced-learn>=0.10.1 18 | shap>=0.36.0 19 | scikit-learn>=0.24,<=1.5.2 20 | lightgbm>=3.0.0 21 | networkx>=3.1 22 | pandas-dq>=1.29 23 | -------------------------------------------------------------------------------- /setup.cfg: -------------------------------------------------------------------------------- 1 | [egg_info] 2 | tag_build = 3 | tag_date = 0 4 | 5 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | import setuptools 2 | 3 | with open("README.md", "r", encoding="utf-8") as fh: 4 | long_description = fh.read() 5 | 6 | setuptools.setup( 7 | name="autoviml", 8 | version="0.2.02", 9 | author="Ram Seshadri", 10 | # author_email="author@example.com", 11 | description="Automatically Build Variant Interpretable ML models fast - now with CatBoost!", 12 | long_description=long_description, 13 | long_description_content_type="text/markdown", 14 | license='Apache License 2.0', 15 | url="https://github.com/AutoViML/Auto_ViML", 16 | packages=setuptools.find_packages(exclude=("tests",)), 17 | install_requires=[ 18 | "numpy<2", 19 | "pandas>=2.0", 20 | "scipy==1.11.4", 21 | "xlrd", 22 | "matplotlib>3.7.4", 23 | "beautifulsoup4", 24 | "emoji", 25 | "ipython", 26 | "jupyter", 27 | "seaborn", 28 | "catboost>=1.2.7", 29 | "textblob", 30 | "nltk", 31 | "regex", 32 | "xgboost>=0.82,<=1.7.6", 33 | "vaderSentiment", 34 | "imbalanced-learn>=0.10.1", 35 | "shap>=0.36.0", 36 | "scikit-learn>=0.24,<=1.5.2", 37 | "lightgbm>=3.0.0", 38 | "networkx>=3.1", 39 | "pandas-dq>=1.29", 40 | ], 41 | classifiers=[ 42 | "Programming Language :: Python :: 3", 43 | "Operating System :: OS Independent", 44 | ], 45 | ) 46 | -------------------------------------------------------------------------------- /sulov_xgboost.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AutoViML/Auto_ViML/0761412417b6dac0683425c372fd7e4f8942d815/sulov_xgboost.png --------------------------------------------------------------------------------