GraphGuard: Contrastive Self-Supervised Learning for Credit-Card Fraud Detection in Multi-Relational Dynamic Graphs

TL;DR

Credit-card fraud detection faces label scarcity and concept drift in streaming transactions. The authors propose GraphGuard, a contrastive self-supervised framework on dynamic multi-relational graphs, leveraging time-aware subgraph sampling and GNN-based learning to detect anomalies without labeled data. The method constructs per-batch graphs, samples target-local subgraphs, and applies a discriminative contrastive objective to produce anomaly scores, evaluated on real and synthetic datasets with insights into when relational information helps. This work demonstrates feasibility of self-supervised graph representations for fraud detection and points to future directions in supervised fine-tuning and richer relational modeling for production FDS deployments.

Abstract

Credit card fraud has significant implications at both an individual and societal level, making effective prevention essential. Current methods rely heavily on feature engineering and labeled information, both of which have significant limitations. In this work, we present GraphGuard, a novel contrastive self-supervised graph-based framework for detecting fraudulent credit card transactions. We conduct experiments on a real-world dataset and a synthetic dataset. Our results provide a promising initial direction for exploring the effectiveness of graph-based self-supervised approaches for credit card fraud detection.

Figures (1)

• Figure 1: Schematic of GraphGuard, inspired by liu_anomaly_2022_COLA. During Graph Construction, each batch $B_i$ comprising daily transactions is transformed into a graph, incorporating data from a sliding window encompassing the most recent $\theta$ days. In the Instance Pair Sampling phase, contrastive instance pairs are sampled for the transaction nodes of day $i$, utilizing preceding transactions as a historical backdrop. Lastly, GNN-Based Contrastive Learning processes the target node and its local subgraph through a GNN to yield node embeddings. The node embeddings of the local subgraph are additionally put through a readout function to summarize their embeddings in an embedding vector. The embedding vector of the subgraph and the target node are then given to a discriminator module, which determines the fit between the target node and the local subgraph.