It's common today to learn predictive user representations for multiple downstream tasks. But these representations can generate unfair predictions. So, in our recent #NeurIPS25 paper, we propose bounding the unfairness across all models for any task with fine-grained control.
