基于谱分析的解释质量结构探索

随着 machine learning models being increasingly considered for high-stakes domains， effective explanation methods 至关重要 to ensure that their prediction strategies are transparent to the user。多年来，已提出 numerous metrics to assess quality of explanations。然而，其 practical applicability 仍不清晰，尤其是 due to a limited understanding of which specific aspects each metric rewards。本文提出 a new framework based on spectral analysis of explanation outcomes，系统地 capture the multifaceted properties of different explanation techniques。我们的分析揭示了两种 distinct factors of explanation quality—stability and target sensitivity—通过 spectral decomposition 直接可见。MNIST 和 ImageNet 上的实验表明，流行的 evaluation techniques（如 pixel-flipping, entropy）部分 capture 这些 factors 之间的 trade-offs。总体而言，我们的 framework 提供了 understanding explanation quality 的基础，指导 develop more reliable techniques for evaluating explanations。