This repository contains the materials for my session 'Explainable ML: Application of different approaches' at the ODSC, Europe conference of September 2020.
Have a look at the requirements.txt to see versions of which packages you may need in order to replicate the code.
You may find it better to work in a virtual environment and install the packages there, instead of potentially distrupting other existing versions you may have of some of these packages. You may choose to work with a pip virtualenv, conda or to avoid the Shell, you can set it up directly in Anaconda navigator.
The powerpoint document contains the slides, which give a definition of explainability and arguments for(and against) it. It also presents a taxonomy of XAI approaches, and contains short technical introductions of the methods applied.
Finally, the jupyter notebook file, XAI_dibates_df.ipynb contains a detailed walk through of a few explainability approaches, using the classical scikit-learn diabetes use case. Though a toy problem, it allows us to demonstrate the following approaches:
- RuleFit
- Partial dependency plots
- Individual conditional expectations
- Global surrogate models
- LIME
- Shap