StreamFair: Online Fair Adaptation Under Temporal Intersectional Drift

Md Fahim Sikder, Ericka Johnson, Fredrik Heintz

Keywords: intersectional fairness, distribution shift, online learning, equalized odds

Abstract: A model that satisfies equalized odds on its training data does not stay fair once the input distribution starts to change, and the gap is the largest on intersectional subgroups whose sample counts may become smaller over time as the stream goes on. We can either let the equalized odds (EO) gap grow over time by doing nothing or apply online learning with cross-entropy which in turn can make the gap worse because the gradient is dominated by the majority intersection group. We present \emph{StreamFair}, which tackles this issue by keeping the deployed model frozen, then attaching a small residual module on top of the frozen model, and training the residual module only when a fairness-aware drift detector detects a shift. The detector observes the worst-intersection EO gap with a cumulative sum and exponentially weighted moving average measurement. Once the detector is initiated and detects drift, the residual module is trained with cross-entropy along with per-intersection penalties on the true-positive rate and false-positive rate gaps and a supervised contrastive term that keeps rate estimates stable on small batches. Our experiments on the ACSIncome dataset spanning 2014 to 2019 in California, Texas and New York show that, \emph{StreamFair} achieves same level fairness and accuracy compared to the always-adapt fair baseline while running $6$ to $10$x fewer updates.

Presentation Format: We prefer to present our paper in a 15-20 minute talk

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

LLM Policy: LLMs may be used to understand the paper and related work but may not be used to assess quality, identify strengths/weakesses, draft the review structure or write the review itself

Submission Number: 34