Feature Importance
A measure of the contribution of each input feature to a model's predictions, either globally (across all predictions) or locally (for a specific prediction).
Feature importance is the most fundamental concept in machine learning explainability — quantifying which input variables a model relies on most when making predictions.
Types of feature importance:
Global feature importance — Aggregates contributions across many or all predictions to describe the model's overall behaviour. Useful for model documentation, bias detection, and regulatory filings. Methods: SHAP summary plots, permutation importance, tree-based impurity.
Local feature importance — Describes which features drove a specific individual prediction. Used for adverse action notices, customer explanations, and debugging edge cases. Methods: SHAP values, LIME coefficients.
Feature importance methods compared:
| Method | Type | Consistent | Model-agnostic | |--------|------|-----------|----------------| | SHAP | Both | Yes | Yes | | LIME | Local | No (sampling) | Yes | | Permutation | Global | Yes | Yes | | Tree impurity (Gini) | Global | No | No (tree-only) |
Feature importance drift: Tracking whether feature importance rankings change over time is a leading indicator of model drift. If a model's top 5 features shift significantly without a corresponding business rationale, this suggests the model is compensating for input distribution changes — an early warning signal captured by Claris's explanation monitoring.