A Generalized Approach to Exploring Subgroup Divergence

As the adoption of machine learning models and automated decision-making systems grows, understanding their behavior in different data subgroups is critical to ensuring their reliability and fairness. Current approaches predominantly focus on characterizing classification performance within subgroups or identifying subgroups where a model performs poorly. However, disparities in the subgroup behavior arise from various sources and encompass a wider range of models and tasks beyond classification performance. There is a need to explore and comprehend subgroup behavior in a broader context, including applications related to ranking, regression models, scoring functions, and studying specific data attributes. We introduce the notion of divergence as a general measure of subgroup anomaly in model or data behavior. Divergence quantifies the difference between a subgroup's performance statistics and the overall dataset statistics. The proposed approach explores all possible subgroups, using frequent pattern-mining techniques to analyze their divergence efficiently. The approach is target-agnostic and model-agnostic, making it suitable for various applications, including classification, regression, and ranking systems. Experimental results on real-world datasets demonstrate the effectiveness and efficiency of the proposed approach in uncovering critical subgroup behaviors.