Rethinking Feature Importance-Evaluating SHAP & TreeSHAP for Tree-Based MLM
summary
Compares SHAP’s assumptions and practical limits with concrete experiments on XGBoost LightGBM CatBoost models, showing kernel SHAP’s sensitivity to label encoding and class imbalance and TreeSHAP’s exact, much faster computation; kernel SHAP can take 2 hours vs 5 minutes for TreeSHAP on the same tree ensemble, and TreeSHAP yields tighter clustering and greater run-to-run stability across IBM HR (1k rows), Kaggle credit (150k), and UCI Adult (48k) datasets, while high‑cardinality categoricals and preprocessing still produce misleading attributions so kernel SHAP vs TreeSHAP requires careful preprocessing.
AI-written summary · about
chapters (13)
- 0:00Feature importance and SHAP
- 3:38SHAP with Minions example
- 6:53SHAP assumptions and limits
- 8:59TreeSHAP for trees
- 9:44Tree-based ML models
- 11:08TreeSHAP exact SHAP values
- 13:11HR workforce case study
- 17:24SHAP vs TreeSHAP on HR data
- 21:36Finance credit scoring case
- 23:42Income data social science
- 26:11KernelSHAP vs TreeSHAP
- 27:35Model interpretability takeaways
- 29:00Q&A on TreeSHAP stability
related talks
31mBeyond the Black Box: Interpreting ML Models with SHAP
35mSHAPtivating Insights: unravelling blackbox AI models
27mInterpretability and explainability in machine learning
35mTaking a peek under the hood - Interpreting black box model
30mTraditional & Novel Feature Selection Approaches
26mShparkley: Scaling Shapley values with Spark
33mBringing life and motion to AI Explainability
36mSHAP beyond the standard graphics: co-design of ML-models in earth sciences