Package to understand Supervised ML Models. This package has been tested with Scikit-Learn and XGBoost library. It should work with any machine learning library that has a predict and predict_proba methods for regression and classification estimators.
There are currently two main sets of capabilities of this package. The first is around understanding "black-box" models via the "Model X-Ray". The second is for probability calibration.
For understanding black-box models, the main entry point is the ModelXRay class. Instantiate it with the model and data. The data can be what the model was trained with, but intended to be used for out of bag or test data to see how the model performs when one feature is changed, holding everything else constant.
For probability calibration, the main class is the SplineCalibratedClassifierCV. Using this class you can train your base model, and the corrective calibration function with just a couple of lines of code. See the examples by following the link below.
Python 2.7 and 3.4+
We have tested this tool to the best of our ability, but understand that it may have bugs. It was developed on Python 3.5, so should work better with Python 3 than 2. Use at your own risk, but feel free to report any bugs to our github. <https://github.com/numeristical/introspective>
$ pip install ml_insights>>> import ml_insights as mli
>>> xray = mli.ModelXRay(model, data)>>> rfm = RandomForestClassifier(n_estimators = 500, class_weight='balanced_subsample')
>>> rfm_cv = mli.SplineCalibratedClassifierCV(rfm)
>>> rfm_cv.fit(X_train,y_train)
>>> test_res_calib_cv = rfm_cv.predict_proba(X_test)[:,1]
>>> log_loss(y_test,test_res_calib_cv)Find the latest version on github: https://github.com/numeristical/introspective
Feel free to fork and contribute!
Free software: MIT license
- Brian Lucena
- Ramesh Sampath
Alex Goldstein, Adam Kapelner, Justin Bleich, and Emil Pitkin. 2014. Peeking Inside the Black Box: Visualizing Statistical Learning With Plots of Individual Conditional Expectation. Journal of Computational and Graphical Statistics (March 2014)