TaylorPODA: A Taylor Expansion-Based Method to Improve Post-Hoc Attributions for Opaque Models

Existing explanation methods lack principled foundations: The paper identifies that many widely-used post-hoc attribution methods rely on "heuristic or only partially justified attribution mechanisms," meaning popular AI explanation tools may not have solid theoretical grounding for how they assign importance to different features.

There's a fundamental tension between principled explanations and practical usefulness: TaylorPODA reveals that attribution methods face a trade-off between being theoretically sound and being adaptable to what users actually need, suggesting that current methods may sacrifice rigor to be useful or vice versa.

A new framework offers more transparent feature attribution: The proposed TaylorPODA method grounds explanations in Taylor expansion mathematics and introduces formal "postulates" (principled requirements) for how to fairly and completely assign model behavior to individual features, providing a more systematic approach than existing heuristics.

The method works across different types of models: Because TaylorPODA admits a "Harsanyi-dividend interpretation," it can explain models that aren't mathematically differentiable, making it more broadly applicable than methods requiring specific mathematical properties—important for explaining diverse AI systems in practice.

Balancing rigor with adaptability is possible: TaylorPODA introduces a "controllable allocation mechanism" that allows explanations to adapt to specific downstream objectives while still maintaining theoretical soundness, suggesting that principled and practical explanations aren't necessarily incompatible.