A Controlled Disentanglement Protocol for Label-Efficient Fraud Detection on Drifting Transaction Graphs

Label-efficient fraud detection on drifting transaction graphs has two design axes: the predictor (feature representation and model class) and the acquisition rule that chooses which cases analysts label. Active learning evaluations typically vary only the acquisition rule, so reported gains can conflate the two. We propose a controlled disentanglement protocol that separates them: a prequential closed loop with a random-acquisition floor, a full cross of predictor against acquisition rule, and pre-specified credibility criteria. Applied across three datasets, two graph constructions, and a ladder of acquisition strategies including Drift-Conditioned Active Learning (DCAL), the protocol locates the dominant lever in the predictor: tree ensembles on raw features exceed a linear head on frozen graph embeddings by +0.26 in area under the precision–recall curve (AUC-PR), a gap localized to the interaction of raw features with a nonlinear model; a single raw-feature random forest (0.704) exceeds the full committee (0.614). Once the predictor is controlled, no acquisition strategy holds a stable advantage beyond a small edge for uncertainty sampling that appears under the linear head and vanishes under the committee predictor. The effect replicates on Elliptic++ transactions (+0.173). Practitioners should invest in representation and model class, and use drift detection as a retraining trigger rather than a labeling rule.