Bias-aware fraud model training using control-group labels and SSL representations

Because only some transactions get checked, the labeled examples can be skewed. This workflow uses a random control sample for fairer labels and combines it with broad self-supervised learning on all traffic.