Local Feature Importance

Unlike global methods that describe average model behavior, local importance explains a single prediction. The attribution values are computed for one input vector at a time, revealing which features pushed that particular prediction higher or lower relative to the baseline. This is critical for auditing high-stakes decisions like loan denials or medical diagnoses, where the reasoning for each case must be individually justified.

Local feature importance methods express a prediction as a linear sum of feature contributions. The final prediction equals the baseline value plus the sum of all individual feature attributions. This additive property ensures that the explanation is complete and self-consistent—every unit of the prediction is accounted for by a specific feature's influence, leaving no unexplained residual.

Every local explanation is inherently contrastive: it explains why the prediction differs from a reference or baseline value. The baseline represents the expected model output when no feature information is known. Feature attributions then quantify how each feature moved the prediction away from this neutral starting point, making explanations relative and context-dependent.

Local importance values carry both magnitude and sign. A positive attribution indicates the feature increased the prediction relative to the baseline, while a negative attribution shows it decreased the prediction. This directional information is essential for understanding not just which features matter, but how they influence the outcome—enabling precise debugging and actionable recourse.

Beyond individual feature contributions, local explanations can capture pairwise interactions where the combined effect of two features differs from the sum of their individual effects. For example, age and income might interact non-additively in a credit risk model. Interaction values distribute credit among feature pairs, revealing synergistic or antagonistic relationships that single-feature attributions miss.

Local feature importance frameworks like SHAP and LIME are designed to be model-agnostic, operating on any black-box model regardless of its internal architecture. They treat the model as a function from inputs to outputs, requiring only the ability to query predictions. This universality makes local explanations a portable auditing tool across diverse model types—from gradient-boosted trees to deep neural networks.