Contrastive Subgroups: Discovering Where Two Populations Differ, and Why

ICLR 2026 Conference Submission21057 Authors

Keywords: Interpretability;Subgroups;

TL;DR: We introduce the concept of contrastive subgroups, describing a subset of indviduals who, despite sharing similar characteristics, exhibit significant differences in a target outcome based on group membership.

Abstract:

Given data from two distinct populations, a contrastive subgroup describes a subset of individuals from both groups who, despite sharing similar characteristics, exhibit significant differences in a target outcome. For example, we want to identify subsets of patients who respond differently to a treatment compared to a control group, or uncover disparities between protected and unprotected groups in fairness analysis. In this work, we formalize the notion of contrastive subgroups and propose a general optimization objective to discover them. To make these discovered subgroups actionable, we provide conditions under which the discovered subgroups allow to make causal inferences. We introduce Subcon, a gradient-based method to discover contrastive subgroups and evaluate it on both synthetic and real-world datasets. The results confirm that our method effectively identifies subgroups that expose significant, informative differences in
real-world datasets.

Primary Area: interpretability and explainable AI

Submission Number: 21057