IMEX Explains Black-Box Model Predictions with Feature Interactions.

Explaining why a black-box predictive model arrives at a specific output is crucial for trust and validation, especially in critical applications. The IMEX framework offers a new method for achieving this by focusing on both individual feature contributions and the complex interactions between features. It moves beyond simple feature importance to uncover how combinations of variables influence predictions. IMEX utilizes two metrics: Static Correlation Power (PCS) for individual feature impact and Interaction Correlation Power (PCI) for non-additive effects. This allows for the analysis of interactions involving more than two features. Experimental validation against existing methods on synthetic datasets demonstrates IMEX's ability to accurately identify relevant feature-level structures, even in the presence of non-linear and multicollinear relationships.