SHAPtivating Insights: unravelling blackbox AI models
summary
SHAP analysis on a tabular bank transaction classifier (XGBoost) exposes concrete findings: a single feature creates 100% accuracy due to label leakage, that feature is removed and the model is retrained, global and local SHAP plots reveal which features truly drive predictions, SHAP guides feature selection and sensitivity checks, and a practical pipeline converts SHAP scores into textual translation for analysts (Excel/text) so alerts become interpretable rather than opaque global and local SHAP values enabling operational use of a XGBoost model.
AI-written summary · about
chapters (22)
- 0:00Introduction
- 1:04Explainable AI purpose
- 4:13Why SHAP for explainability
- 6:21Classifier with SHAP in Python
- 8:30SHAP values for insights
- 9:36Global feature importance
- 12:07Feature sensitivity plots
- 13:58Decision plots analysis
- 15:24Local predictions explainability
- 16:27Waterfall plots breakdown
- 17:53Force plots contributions
- 18:55Translating SHAP insights
- 22:00Key takeaways and limits
- 23:22SHAP real-world impact
- 24:27Q&A: Multiclass models
- 25:33Q&A: SHAP data dependency
- 27:41Q&A: Local vs global SHAP
- 29:24Q&A: Edge cases handling
- 30:08Q&A: Model restrictions
- 31:17Q&A: SHAP explainability illusion
- 32:42Q&A: EU AI Act depth
- 34:27Closing remarks
related talks
31mBeyond the Black Box: Interpreting ML Models with SHAP
35mTaking a peek under the hood - Interpreting black box model
27mInterpretability and explainability in machine learning
29mOpeningTheBlackBox
39mExplainable AI as a Conventional Data Analysis Tool
01h 36mOpen the Black Box: an Introduction to Model Interpretability with LIME and SHAP
33mBringing life and motion to AI Explainability
30mTrustworthy AI - opening up the black box