Provides model-agnostic interpretability via DALEX and SHAP-like methods. Essential for understanding which features drive predictions and for generating hypotheses in biomarker discovery.
Details
## Why Interpretability Matters
For biomarker discovery, understanding WHY a model makes predictions is as important as accuracy. Key use cases: - Identify which genes/miRNAs drive classification - Detect potential confounders or batch effects - Generate hypotheses for biological validation - Build trust for clinical adoption
## Methods Provided
- **Feature Importance**: Permutation-based importance (model-agnostic) - **Partial Dependence**: How features affect predictions marginally - **SHAP Values**: Additive feature attributions (via iml or approximation) - **Correlation Warnings**: Flag high-correlation features that may mislead
## Fold-Aware Design
All interpretability methods respect the cross-validation structure: - Explanations are computed per-fold using training data only - Aggregated explanations show stability across folds - Unstable features are flagged