PREF-Gate: Provenance-Constrained Relational Evidence Fusion with Validation-Gated Selection for Graph Fraud Detection

Graph fraud detection often relies on two distinct information sources: label-free graph context (e.g., node attributes and structural patterns) and label-derived neighborhood evidence (e.g., fraud rates of neighbors). A critical, often overlooked issue is evidence provenance—ensuring that neighborhood risk is not contaminated by validation or test labels. The authors introduce PREF-Gate, a decision framework that enforces a strict label-provenance contract. It utilizes two fixed experts: a context expert that operates solely on label-free features and an evidence expert that incorporates empirical-Bayes summaries of training-label-only neighborhood risk. A finite validation gate then selects the best-performing expert or a pre-specified probability mixture before any test inference occurs.

Across three benchmark datasets (Amazon, YelpChi, and TFinance), PREF-Gate demonstrates competitive performance, achieving mean AUPRC values of 0.9085, 0.8104, and 0.913, respectively. Notably, the gate's behavior varies by dataset: it consistently selects the label-free context expert for Amazon and YelpChi, rejecting the label-derived evidence, while it selects a mixture of both experts for TFinance. This result highlights that label-derived relational evidence is not universally beneficial and should only be admitted when validation performance explicitly justifies its inclusion.

This paper addresses the "black box" nature of many graph neural network pipelines by providing an auditable decision framework. By separating evidence construction from decision-making and enforcing a strict information boundary, PREF-Gate prevents data leakage and allows researchers to account for failure modes. It demonstrates that adding more complex, label-dependent features does not always improve model performance and can sometimes introduce unnecessary variance, suggesting that simpler, label-free context models are often more robust for operational fraud detection.