Explainable ML Course Materials & Projects

This repository contains my course materials and projects for the Explainable Machine Learning course at UW-Madison taught by Prof. Kris Sankaran. It is forked from the official course repository

Projects: Explainability of LightGCN-based Recommender Systems

• LightGCN+Integrated_Gradients:
  This project applies Integrated Gradients to a feature-augmented LightGCN trained on the MovieLens 1M dataset to attribute recommendation scores to user demographics and movie genres. It demonstrates how gradient-based attribution can explain individual recommendations while also revealing limitations related to feature correlation and non-causal interpretations.

  • Report: stat992hw1.pdf
  • Experiments: Explain_Recommender.ipynb

• LightGCN+CAV:
  This project explores Concept Activation Vectors (CAVs) for concept-level interpretability in embedding-based LightGCN recommenders. Movie genres are treated as semantic concepts, enabling efficient analysis of genre geometry in the latent space and user-specific conceptual sensitivity, and further supporting concept-driven customization of recommendations through interpretable embedding manipulation.

  • Reports: stat992hw2.pdf, stat992hw3.pdf
  • Experiments: Explain_LightGCN.ipynb, hw3_explore.ipynb
  • Model&Method: LightGCN.py, RecommenderCAV.py, MovieDataProcessor.py, RecommenderTrainer.py

Course Materials

• Mynotes: Personal weekly notes and reflections on course topics
• notes, demos, exercises, logistics: Official course materials including lecture notes, code demonstrations, assignments, and syllabus