To assess the robustness of FAs in the presence of collinear or redundant features, [Salih et al. (2022)] introduce a metric that evaluates the stability of feature rankings as the most important features are iteratively removed and the model is retrained. This procedure mirrors the setup of the Remove and Retrain paradigm (Metric “Retrained Model Evaluation”), but shifts the focus from prediction performance to the behavior of the explanation itself.
After each retraining step, the explanans is recomputed, and the ranks of the remaining features are compared to the previous ones. A large shift in ranking indicates that redundant or weakly relevant features have taken over the role of more informative ones, suggesting that the attribution method lacks robustness in the face of redundancy and may not reflect the true rationale behind the model's prediction.
To address this issue, [Salih et al. (2024)] propose the Modified Informative Position (MIP), which aims to stabilize the interpretation by providing a more resilient ranking structure across iterative retraining steps.