Directional Influence Function for Constrained Learning Models

A novel approach known as the Directional Influence Function (DIF) has been introduced to assess how training data impacts models subjected to constraints like fairness, safety, robustness, regularization, and physics or logic limitations. Traditional influence functions falter in constrained environments, as data changes can modify both the objective and the feasible region, resulting in estimates that breach feasibility. By integrating these constraints into the influence estimation process through the optimality conditions of constrained learning, DIF seeks to enhance the interpretability and robustness of constrained models. This research is detailed in a paper available on arXiv under ID 2607.23388.