Calibrated Data-Dependent Constraints with Exact Satisfaction Guarantees
This addresses fairness and stability issues in machine learning models, but it is incremental as it builds on existing constraint reformulation methods.
The paper tackles the problem of training machine learning models with data-dependent constraints, such as fairness or stability goals, by reformulating them to be calibrated, ensuring exact satisfaction of expected value constraints with a user-prescribed probability. It demonstrates efficacy on a fairness-sensitive classification task, guaranteeing fairness at test time.
We consider the task of training machine learning models with data-dependent constraints. Such constraints often arise as empirical versions of expected value constraints that enforce fairness or stability goals. We reformulate data-dependent constraints so that they are calibrated: enforcing the reformulated constraints guarantees that their expected value counterparts are satisfied with a user-prescribed probability. The resulting optimization problem is amendable to standard stochastic optimization algorithms, and we demonstrate the efficacy of our method on a fairness-sensitive classification task where we wish to guarantee the classifier's fairness (at test time).