Exploring Feature Redundancy and Synergy with FACET 2.0
summary
FACET 2.0 implements SHAP vector projection to turn per-instance SHAP values into feature vectors, then uses angles and an orthogonal decomposition to quantify redundancy and synergy, not just correlation. It is open source on the linked repo and runs on real problems with thousands of variables though typical workflows narrow to 20–50 candidates. Examples show dropping redundant features preserves performance and maintains R² ≈ 0.92–0.93, and a previously split credit for a feature doubles after removal.
AI-written summary · about
chapters (14)
- 0:00FACET 2.0 introduction
- 0:28Model explanations matter
- 1:55Synergy and redundancy in features
- 4:50FACET feature interactions
- 8:44Synergy and redundancy matrices
- 10:57FACET workflow demo
- 15:37SHAP values and importance
- 16:41Redundancy in dataset
- 17:46Feature synergy analysis
- 18:30Feature clustering dendrograms
- 19:32Pipeline after redundancy removal
- 21:38Feature impact simulation
- 24:49FACET installation and performance
- 25:52Q&A scaling and use
related talks
35mTaking a peek under the hood - Interpreting black box model
27mFast, intuitive feature selection via regression on Shapley values
27mIs This Feature Interesting? Interfeat: AI for Insight Discovery
01h 28mFeature Engineering Made Simple
28mCifar-10 Exploratory Data Analysis
42mMitigating Bias in Models with SHAP and Fairlearn
40mA Practical Guide to Dimensionality Reduction Techniques
31mBeyond the Black Box: Interpreting ML Models with SHAP