Attributions Beyond Neural Networks: The Linear Program Case

Linear Programs Need Better Explainability Tools: The paper addresses a significant gap in responsible AI by showing that while we can solve high-dimensional linear programs (mathematical optimization problems used across industries), we currently lack good methods to understand why these solutions were chosen—a critical need for trustworthy decision-making systems.

Attribution Methods Produce Similar Results in Many Cases: The research finds that two popular explainability techniques (Saliency and LIME) "reveal indistinguishable results up to perturbation levels" when applied to linear programs, suggesting that organizations may get comparable insights from either method, though the choice matters for certain applications.

"Directedness" Matters for Choosing Explanation Methods: The paper identifies a key property called "Directedness"—whether an attribution method explains how features affect outcomes when they increase versus decrease—as the main factor distinguishing different explanation approaches, helping practitioners select the right tool for their specific transparency needs.

Baseline Selection Remains Unsolved: The research highlights that "the baseline selection problem" for Integrated Gradients (a popular explanation technique) extends beyond computer vision into optimization problems, meaning practitioners still lack clear guidance on how to properly configure this method across different domains.

Neural Network Explanation Methods Can Extend Beyond Neural Networks: By encoding linear programs as neural networks, the authors demonstrate that explainability techniques developed for deep learning can be adapted to explain other types of AI systems, potentially broadening the toolkit for responsible AI across different algorithmic approaches.