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

Jul 16, 2026

Relational fraud detection can leverage label-free graph context and label-derived neighborhood evidence, yet these sources follow different validity conditions. Neighborhood risk becomes invalid when a queried node’s label, or any validation/test label, enters its construction. PREF-Gate formulates provenance-constrained relational evidence use as an auditable framework with two fixed experts and a finite validation gate. The context expert stays label-free; the evidence expert adds training-label-only neighborhood risk with empirical-Bayes uncertainty summaries and a selection policy tied to validation.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages40

Reading Time1 h 41 min

Abstract

Introduction

Evidence provenance

Evidence reliability

PREF-Gate framework

Model identity

Evidence and decision workflow

arXiv :2607 . 1 12 12v 1 [ cs .AI] 13 Jul 2026 PREF-Gate: Provenance-Constrained Relational Evidence Fusion with Validation-Gated Selection for Graph Fraud Detection Liming Liua,∗, Chao Hua , Mingfei Lub , Yiwei Gea , Xingle Lia , Heyuan Shia a Central South University, Changsha, China b University of Technology Sydney, Australian Artificial Intelligence Institute, Sydney, Australia Abstract Relational fraud detection can exploit both label-free graph context and label-derived neighborhood evidence, but these two information sources obey different validity conditions. In particular, neighborhood risk becomes invalid when a queried node’s own label, or any validation or test label, enters its construction. We formulate this issue as provenance-constrained relational evidence use and present PREF-Gate, an auditable decision framework with two fixed experts and a finite validation gate. The context expert uses attributes, one-hop means, feature residuals, and degree descriptors without labels. The evidence expert adds self-excluded, training-label-only neighborhood risk and empirical-Bayes summaries that expose support, uncertainty, availability, and shrinkage. Before test inference, the gate selects either expert or one of three pre-specified probability mixtures and fixes the decision threshold. On Amazon, YelpChi, and TFinance, using five identical stratified splits and 14 same-protocol methods, PREF-Gate obtains mean AUPRC values of 0 .9085, 0 .8104, and 0 .8913. It selects the label-free expert on all Amazon and YelpChi splits and an evidence mixture on all TFinance splits. Thus, the main result is conditional rather than universal: label-derived relational evidence is useful only where held-out validation supports it. The framework couples competitive ranking performance with an explicit labelprovenance contract, finite selection policy, failure accounting, and review- ∗ Corresponding author Email addresses: [244512048@cs.edu.cn](244512048@cs.edu.cn) (Liming Liu), [8206240605@csu.edu.cn](8206240605@csu.edu.cn) (Yiwei Ge), [8206240606@csu.edu.cn](8206240606@csu.edu.cn) (Xingle Li) budget evaluation, providing an auditable knowledge-based decision pipeline for graph fraud detection. Keywords: graph fraud detection, relational evidence, label provenance, validation-gated fusion, empirical Bayes, decision support 1. Introduction Fraud detection is an imbalanced decision problem embedded in a relational system. A platform may inspect only a small fraction of users, reviews, or transactions, while fraudulent entities share devices, products, counterparties, and structural patterns. These relations provide contextual knowledge that is not present in an entity’s attributes. Graph neural networks (GNNs) offer one way to learn from such dependencies [1, 2, 3, 4], but model expressiveness alone does not determine whether relational information is valid or useful for an operational decision. The first unresolved issue is evidence provenance. One-hop feature context is label free, whereas neighborhood fraud risk is knowledge derived from observed labels. The latter can be highly predictive while being scientifically invalid if validation or test labels enter its numerator, denominator, prior, or selection rule. Even a training node can leak its target through a self-loop or a global prior. Because transductive graph pipelines commonly load all nodes and edges together, the distinction must be enforced at the evidence operator, not left as an informal property of the data split. The second issue is evidence reliability. A neighborhood rate supported by many eligible labels is different from the same rate supported by one label. On another dataset, raw attributes and label-free neighborhood summaries may already be sufficiently discriminative, so adding label-derived evidence can reduce generalization. A fixed fusion policy can therefore turn an occasionally useful signal into a systematic source of variance.

FAQs

What makes neighborhood evidence invalid in graph fraud detection under PREF-Gate?

Neighborhood risk becomes invalid if the construction uses the queried node’s own label or any validation/test labels, including through terms like numerator/denominator/prior or selection rules.

How do PREF-Gate’s two experts differ?

The label-free context expert uses attributes, one-hop means, feature residuals, degree descriptors, and an isolation indicator. The evidence expert augments this with training-label-only neighborhood risk and empirical-Bayes summaries (mean, variance, support, availability, shrinkage).

How does the validation gate influence the final decision at test time?

Before requesting test probabilities, the gate selects either the label-free expert or a training-label evidence mixture among pre-specified candidates using validation metrics, then freezes the decision threshold and applies it to test inference.

Jul 16, 2026

Upload by Theodora