Interaction-Aware Influence Functions for Group Attribution
A new arXiv paper (2605.15675v1) introduces interaction-aware influence functions that capture how groups of training examples jointly affect a target function, such as held-out loss. Standard influence functions sum individual influences, missing redundancy or complementarity between examples. The proposed method expands the target to second order around trained parameters, adding a pairwise interaction term. Empirical evaluation on six dataset-model pairs (including logistic regression, MLPs, and ResNet-9) shows improved attribution.
Key facts
- arXiv paper 2605.15675v1 proposes interaction-aware influence functions.
- Standard influence functions sum individual influences of group members.
- Sum does not capture joint effects like redundancy or complementarity.
- New method expands target to second order around trained parameters.
- Estimator augments standard sum with pairwise interaction term.
- Evaluated on six dataset-model pairs including logistic regression, MLPs, and ResNet-9.
Entities
Institutions
- arXiv