GRAD-E1394 Deep Learning - Assignment 3
This tutorial introduces three popular post-hoc, model-agnostic explanability methods for "black-box" machine-learning models: LIME, SHAP, and Diverse Counterfactual Explanations (DiCE), as well as their implementation in Python.
We first train a basic Feed-Forward-Neural-Network classifier on the German Credit Data set. By introducing synthetic test cases, based on real life scenarios, we show that the model discriminates one person over the other, although intuitively one would expect a different outcome. We then apply LIME, DiCE and SHAP and interpret each method's outcomes in the light of our toy example.
In addition to this extensive and technical tutorial, we provide readers with presentation slides, taking a more conceptual angle on the topic.
Authors:
- Benjamin Adoba Ayida (Github)
- Padmavathi Narayana Reddy, (Github)
- Luis Windpassinger (Github)
- Franco Bastida (Github)
