New explainability method IMEX quantifies feature importance and interactions in black-box models

A new explainability method called IMEX (Interaction-Based Model Explanation) has been proposed to address the opacity of black-box predictive models. Unlike traditional approaches that focus solely on individual feature importance, IMEX introduces a framework to quantify both the contributions of individual variables and the significance of interactions among variables in determining model predictions.

The IMEX framework is built around two complementary metrics: Static Correlation Power (PCS) and Interaction Correlation Power (PCI). PCS quantifies the contribution of individual features to the target prediction, while PCI captures non-additive effects arising from interactions among features. This dual-metric approach enables the construction of an interpretability map of predictions, highlighting not only which features matter but also how they interact.

The work experimentally validates the PCS component of IMEX by comparing it with the INVASE method on three synthetic datasets with known underlying structures. These datasets are designed to include non-linear relationships, conditional dependencies, and multicollinearity between input features and prediction targets.

The results indicate that IMEX can recover relevant feature-level structures under these challenging conditions, suggesting potential utility for diagnosing and validating complex predictive models in domains where interpretability is essential.