Influence functions are a standard tool for estimating how removing or reweighting training data would change a model, without retraining. They offer an efficient alternative to full retraining for assessing the effects of individual data points.

For groups of points, the standard first-order estimate takes a particularly simple form: it reduces individual contributions to a sum. A new arXiv preprint, 'Groups Don't Move as One: A Bregman Clustering View of Linearized Group Influence,' focuses on this group setting.

The paper introduces a Bregman clustering view of linearized group influence, as its title indicates. The framing suggests that groups do not move as one under perturbation, and that the summed first-order estimate may not capture the full picture. The preprint is available on arXiv.