Fairness

Long-Term Fairness and ML Safety

Overview

Many ML-enabled systems operate in dynamic environments: the system’s decisions change the environment, and those changes feed back into its future inputs. Certain self-reinforcing loops can amplify errors, entrench bias, and cause fairness violations in the long term even when immediate outcomes are fair. In predictive policing, for example, a model that flags a neighborhood as high-crime sends more patrols there, producing more recorded arrests, which the model reads as even higher crime. The same pattern appears in loan approvals that affect credit scores and in medical risk scoring that influences treatment access.

Fair-AutoML: Performance-Aware Fairness Repair

Overview

Bias mitigation algorithms typically work only in specific situations and often repair fairness at the cost of large accuracy drops, making them impractical for critical decision-making software. We treated fairness repair as a performance-aware optimization problem: fix the bias in a buggy model without ruining its accuracy.

Fair-AutoML, built on a state-of-the-art AutoML tool, made this concrete through two changes to standard AutoML: an optimization function that incorporates fairness objectives alongside accuracy, and a fairness-aware search space over candidate model configurations. A search-space pruning method further reduced computational cost and repair time.

Fairify: Fairness Verification of Neural Networks

Overview

We built Fairify, an SMT-based approach that verifies individual fairness of deep neural networks in production. Individual fairness requires that any two individuals who differ only in protected attributes such as race, sex, or age receive similar predictions; unlike group metrics, it captures worst-case discrimination. The property is hard to verify because it must be checked globally over the input domain and because of the non-linear computation nodes in the network.

Causal Fairness in Machine Learning Pipelines

Overview

Most fairness research treated a machine learning model as a single black box, measuring bias only from its final predictions. Real pipelines, however, contain an ordered set of components — data filtering, imputation, encoding, feature transformation, training, tuning — and each can affect the fairness of the resulting model. We investigated fairness at the component level: using causal reasoning, we intervened on one stage at a time, constructed an alternative pipeline without that stage, and measured the resulting prediction disparity to attribute unfairness to specific components.