TeMP-TraG: Edge-based Temporal Message Passing in Transaction Graphs

TL;DR

TeMP-TraG introduces a temporal, edge-aware extension to multigraph GNNs for transaction graphs to combat financial crime. It combines multigraph message passing with a time-based weighting that prioritizes recent transactions, and can be applied on top of MEGA-GNN or Multi-GNN. Experiments on ETH phishing and IBM AML datasets show $6.19\%$ average improvement over four state-of-the-art GNNs, with edge-timestamp aggregation strategies influencing performance. The approach advances time-sensitive detection in financial networks and suggests future work incorporating geographic data and AML-domain knowledge.

Abstract

Transaction graphs, which represent financial and trade transactions between entities such as bank accounts and companies, can reveal patterns indicative of financial crimes like money laundering and fraud. However, effective detection of such cases requires node and edge classification methods capable of addressing the unique challenges of transaction graphs, including rich edge features, multigraph structures and temporal dynamics. To tackle these challenges, we propose TeMP-TraG, a novel graph neural network mechanism that incorporates temporal dynamics into message passing. TeMP-TraG prioritises more recent transactions when aggregating node messages, enabling better detection of time-sensitive patterns. We demonstrate that TeMP-TraG improves four state-of-the-art graph neural networks by 6.19% on average. Our results highlight TeMP-TraG as an advancement in leveraging transaction graphs to combat financial crime.

Figures (6)

• Figure 1: Comparison of TeMP-TraG's temporal message passing with traditional message passing, showing that edges are weighted based on temporal proximity.
• Figure 2: Overview of TeMP-TraG. Given a transaction graph $\mathcal{G}$, labels $\hat{\mathcal{Y}}$ are generated for its nodes and edges following graph sampling and embedding.
• Figure 3: Multigraph transformation by edge feature and timestamp aggregation. For brevity, we label the edges with an indexed edge and timestamp (e.g., $z_1$, $t_1$) and we use colour encoding for aggregated edge (e.g., $z_o$, $t_o$ for orange edges).
• Figure 4: Illustration of temporal message passing with edge aggregation.
• Figure 5: Performance of different aggregation strategies for edge timestamps in TeMP-TraG (GIN with agg) according to three metrics.

Theorems & Definitions (3)

• definition thmcounterdefinition: Transaction Graph
• definition thmcounterdefinition: Neighbours
• definition thmcounterdefinition: Node Classification