Model Explanation Disparities as a Fairness Diagnostic

In recent years, there has been a flurry of research focusing on the fairness of machine learning models, and in particular on quantifying and eliminating bias against subgroups. One prominent line of work generalizes the notion of subgroups beyond simple discrete classes by introducing the notion of a "rich subgroup," and seeks to train models that are calibrated or equalize error rates with respect to these richer subgroup classes. Largely orthogonally, there has been growing recognition of the importance of understanding how subgroups of the dataset are being treated relative to the rest of the dataset. It can easily be shown that certain training features may be significantly more important (or less important) on a discrete subgroup compared to the whole dataset with this difference being called Feature Importance Disparity (FID). However, there are an exponentially large number of rich subgroups defined by a structured class of functions over protected features (such as race, gender, age, etc.) and there are many ways that feature importance can be defined. In this paper, we develop two approaches to efficiently search the rich subgroup space and find feature/subgroup pairs with large FID that fit within a specified subgroup size. The first approach considers feature importance metrics which are separable and models a two-player, zero-sum game to reduce the computation of subgroups with high FID of constrained size to a cost-sensitive classification problem. The second approach considers non-separable importance metrics and uses heuristic optimization techniques to converge on the subgroups. Both of these approaches were tested on 4 different datasets with multiple importance notions and found feature/subgroup pairs that had high FID, often by orders of magnitude, and yield interesting discussions about the reliability and fairness of the datasets.