hard20 pts · first answer countsretrainingmonitoring

Model affects its own training data

A fraud model blocks transactions it scores as risky. Retraining uses chargebacks as labels. Over time the model seems to get better on paper but misses new fraud patterns. Why?

A.Chargebacks are always wrongB.Retraining too often causes overfitting automaticallyC.Blocked transactions never get labels, so the model creates a feedback loop and learns only from what it lets through; keep a small randomised holdout or exploration to get unbiased labelsD.The model needs a larger batch size

← Hundreds of low-traffic models Loading a model from a public hub →