Summary
This arXiv paper (2610.35827) by Sai Ganesh Nagarajan, Toshinori Yamauchi, and Hiroshi Kera addresses a problem in Shapley-value-based feature attribution for computer vision. Some features, termed contextual distractors, can reduce the target-class score when added to a complete genuine image even while the predicted label stays unchanged. The authors formalize these distractors and study their attributions under semivalues and the least core. Under a weak average marginal monotonicity assumption, they derive quantitative bounds on the semivalue-least-core gap and sufficient conditions based on the probability of sampling smaller contexts. They also prove that shifting semivalue weights toward larger contexts decreases the signed gap. Controlled image-classification experiments and evaluations on Salient ImageNet provide strong empirical support for the theory. This work is relevant for researchers in explainable AI, computer vision, and cooperative game theory applied to machine learning interpretation.
Paper Overview
Field: Computer Vision
Authors: Sai Ganesh Nagarajan, Toshinori Yamauchi, Hiroshi Kera
Published: 2026-10-26
arXiv:
2610.35827Abstract
Shapley-value-based explanations are widely used for feature attribution in machine learning. However, in computer vision, a feature may reduce the target-class score when added to the complete genuine image, even when the predicted label remains unchanged. The authors formalize such features as
contextual distractors and investigate their attributions under semivalues and the least core.
Key Contributions
- Formalization of "contextual distractors" — features whose addition to the complete genuine image lowers the target-class score without changing the predicted label.
- Under a weak average marginal monotonicity assumption, derivation of quantitative bounds on the semivalue–least-core gap.
- Sufficient conditions based on the probability of sampling smaller contexts.
- Proof that shifting semivalue weights toward larger contexts decreases the signed gap.
Empirical Validation
Controlled image-classification experiments and evaluations on Salient ImageNet provide strong empirical support for the theoretical findings.Links
- Paper: https://arxiv.org/abs/2610.35827
---
*Automatically collected on 2026-10-28*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178636543