Skip to contents

Functions to compute SHAP values with warnings about correlated features. High correlations between features can make SHAP values misleading because: 1. Importance can be arbitrarily split between correlated features 2. SHAP values may not reflect true causal relationships 3. Removing one correlated feature could dramatically change others' SHAP values

References

Hooker et al. (2021). Unrestricted Permutation Forces Extrapolation. Lundberg & Lee (2017). SHAP: A Unified Approach to Interpreting Model Predictions.