Beyond the Black Box: Interpreting ML Models with SHAP
summary
Shows how SHAP turns model outputs into feature attributions using XGBoost with TreeExplainer on a 45k-sample Portuguese bank dataset (AUC ~0.79) and 1D CNN and DeepExplainer on a 30-person wearable sensor set (560 features, sequence length 64, six classes). Demonstrates summary and dependency plots, local bar explanations, and practical choices like 100–150 background samples. Calls out costs: exponential exact SHAP, correlation ambiguity, model-only (not causal) explanations.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:49Talk agenda and AI importance
- 3:59SHAP theory and Shapley values
- 8:54SHAP calculation methods
- 10:20Case study: XGBoost trees
- 13:11Global SHAP: bank marketing
- 17:28Dependency plots analysis
- 18:52Case study: time series NN
- 23:50Global SHAP: activity recognition
- 26:39Local SHAP: individual predictions
- 28:07SHAP limitations and cautions
- 30:15Conclusion
related talks
35mTaking a peek under the hood - Interpreting black box model
35mSHAPtivating Insights: unravelling blackbox AI models
27mInterpretability and explainability in machine learning
01h 36mOpen the Black Box: an Introduction to Model Interpretability with LIME and SHAP
39mExplainable AI as a Conventional Data Analysis Tool
34mRethinking Feature Importance-Evaluating SHAP & TreeSHAP for Tree-Based MLM
36mSHAP beyond the standard graphics: co-design of ML-models in earth sciences
44mInterpretable Machine Learning with Python Examples