SHAP Waterfall Breakdown (Local Accuracy)
Actionable Algorithmic Recourse
As DhanushKumar notes: Instead of telling an applicant why a 500-tree ensemble rejected them, counterfactuals answer the human question: “What is the smallest realistic change that flips this outcome?”
The “PDP Trap”: Aggregate Partial Dependence vs. Subgroup ICE
The bold blue line represents average Partial Dependence (PDP). When subgroups interact nonlinearly, the aggregate line appears flat and uninformative, masking the fact that Subgroup A improves while Subgroup B degrades. Individual Conditional Expectation (ICE) reveals this hidden divergence.
LIME Perturbation & Neighborhood Instability Simulator
LIME approximates local boundaries by randomly sampling Gaussian noise around applicant instance x. Because of Monte Carlo variance, re-running LIME with small sample sizes causes coefficient drift. In contrast, SHAP values are mathematically deterministic and unique.
| Feature | Deterministic SHAP φi | LIME Run #1 β | LIME Run #2 (Re-sampled) | Observed Drift (Δ) | Stability Verdict |
|---|
3-Axis Interpretability Taxonomist
Every interpretability method is categorized across three fundamental orthogonal axes: Intrinsic vs. Post-Hoc, Model-Specific vs. Model-Agnostic, and Local vs. Global.
SHAP TreeExplainer
Axes: Post-Hoc • Model-Specific • Local & GlobalExploits decision tree leaf paths to calculate exact polynomial-time Shapley attributions with mathematical additivity guarantees.
LIME (Tabular)
Axes: Post-Hoc • Model-Agnostic • LocalFits weighted sparse linear surrogate in local Gaussian neighborhood. High flexibility; sensitive to sampling randomness.
PDP & ICE
Axes: Post-Hoc • Model-Agnostic • Global (PDP) / Instance (ICE)Marginal effect visualization revealing non-linearities and threshold shifts. Pair PDP with ICE to detect subgroup masking.
Counterfactual Search
Axes: Post-Hoc • Model-Agnostic • Local RecourseSolves optimization for minimum perturbation to flip classification decision. Directly actionable for consumers under EU AI Act.
Logistic Regression
Axes: Intrinsic • Model-Specific • Global & LocalCoefficients directly translate to log-odds changes. Transparent by design, but requires manual feature engineering for interactions.
Permutation Importance
Axes: Post-Hoc • Model-Agnostic • GlobalRandomly shuffles a feature column on validation holdout to quantify performance degradation. Fast, but collinearity can mask duplicate features.