Interaction-Aware Influence Functions for Group Attribution
The paper introduces an interaction-aware influence function designed for group attribution in machine learning. This new approach enhances the standard method by accounting for interactions between examples, which can be either redundant or complementary. Empirical evaluations show that the proposed method outperforms traditional influence functions in various settings.
- ▪The interaction-aware influence function improves the estimation of how groups of examples affect a target function.
- ▪It incorporates a pairwise interaction term that captures the alignment of effects between examples.
- ▪The method was tested on multiple dataset-model pairs and showed better performance than first-order influence methods.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.15675 (cs) [Submitted on 15 May 2026] Title:Interaction-Aware Influence Functions for Group Attribution Authors:Jaeseung Heo, Kyeongheung Yun, Youngbin Choi, Sehyun Hwang, Jungseul Ok, Dongwoo Kim View a PDF of the paper titled Interaction-Aware Influence Functions for Group Attribution, by Jaeseung Heo and 5 other authors View PDF HTML (experimental) Abstract:Influence functions approximate how removing a training example changes a quantity of interest, called the target function, such as a held-out loss. To estimate the influence of a group of examples, the standard practice is to sum the individual influences of its members.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
