**An integrated Python toolbox for interpretable machine learning** `pip install PiML` 🎄 **Dec 1, 2023:** V0.6.0 is released with enhanced data handling and model analytics. :rocket: **May 4, 2023:** V0.5.0 is released together with PiML user guide. :rocket: **October 31, 2022:** V0.4.0 is released with enriched models and enhanced diagnostics. :rocket: **July 26, 2022:** V0.3.0 is released with classic statistical models. :rocket: **June 26, 2022:** V0.2.0 is released with high-code APIs. :loudspeaker: **May 4, 2022:** V0.1.0 is launched with low-code UI/UX.

PiML (or π-ML, /ˈpaɪ·ˈem·ˈel/) is a new Python toolbox for interpretable machine learning model development and validation. Through low-code interface and high-code APIs, PiML supports a growing list of inherently interpretable ML models:

1. GLM: Linear/Logistic Regression with L1 ∨ L2 Regularization
2. GAM: Generalized Additive Models using B-splines
3. Tree: Decision Tree for Classification and Regression
4. FIGS: Fast Interpretable Greedy-Tree Sums (Tan, et al. 2022)
5. XGB1: Extreme Gradient Boosted Trees of Depth 1, with optimal binning (Chen and Guestrin, 2016; Navas-Palencia, 2020)
6. XGB2: Extreme Gradient Boosted Trees of Depth 2, with effect purification (Chen and Guestrin, 2016; Lengerich, et al. 2020)
7. EBM: Explainable Boosting Machine (Nori, et al. 2019; Lou, et al. 2013)
8. GAMI-Net: Generalized Additive Model with Structured Interactions (Yang, Zhang and Sudjianto, 2021)
9. ReLU-DNN: Deep ReLU Networks using Aletheia Unwrapper and Sparsification (Sudjianto, et al. 2020)

PiML also works for arbitrary supervised ML models under regression and binary classification settings. It supports a whole spectrum of outcome testing, including but not limited to the following:

1. Accuracy: popular metrics like MSE, MAE for regression tasks and ACC, AUC, Recall, Precision, F1-score for binary classification tasks.
2. Explainability: post-hoc global explainers (PFI, PDP, ALE) and local explainers (LIME, SHAP).
3. Fairness: disparity test and segmented analysis by integrating the solas-ai package.
4. WeakSpot: identification of weak regions with high residuals by slicing techniques.
5. Overfit: identification of overfitting regions according to train-test performance gap.
6. Reliability: assessment of prediction uncertainty by split conformal prediction techniques.
7. Robustness: evaluation of performance degradation under covariate noise perturbation.
8. Resilience: evaluation of performance degradation under different out-of-distribution scenarios.

Installation | Examples | Usage | Citations

Installation

pip install PiML  

Low-code Examples

Click the ipynb links to run examples in Google Colab:

1. BikeSharing data: ipynb
2. CaliforniaHousing data: ipynb
3. TaiwanCredit data: ipynb
4. Fairness_SimuStudy1 data: ipynb
5. Fairness_SimuStudy2 data: ipynb
6. Upload custom data in two ways: ipynb
7. Deal with external models: ipynb

Begin your own PiML journey with this demo notebook.

High-code Examples

The same examples can also be run by high-code APIs:

1. BikeSharing data: ipynb
2. CaliforniaHousing data: ipynb
3. TaiwanCredit data: ipynb
4. Model saving: ipynb
5. Results return: ipynb

   Low-code Usage on Google Colab

Stage 1: Initialize an experiment, Load and Prepare data

from piml import Experiment
exp = Experiment()

exp.data_loader()

exp.data_summary()

exp.data_prepare()

exp.data_quality()

exp.feature_select()

exp.eda()

Stage 2: Train intepretable models

exp.model_train()

Stage 3. Explain and Interpret

exp.model_explain()

exp.model_interpret() 

Stage 4. Diagnose and Compare

exp.model_diagnose()

exp.model_compare()

exp.model_fairness()

exp.model_fairness_compare()

Arbitrary Black-Box Modeling

For example, train a complex LightGBM with depth 7 and register it to the experiment:

from lightgbm import LGBMClassifier
exp.model_train(LGBMClassifier(max_depth=7), name='LGBM-7')

Then, compare it to inherently interpretable models (e.g. XGB2 and GAMI-Net):

exp.model_compare()

Citations