Temporal-Context Edge Enrichment for Financial Crime Detection

Abstract

Financial crime detection requires identifying rare illicit activity within large transaction networks, where suspicious behaviour can depend on both relationships between accounts and patterns in their transaction histories. Existing graph-based methods capture structural information, yet typically represent temporal behaviour only through transaction attributes or shared account representations. This thesis investigates whether representations from pretrained time-series foundation models (TSFMs) can provide richer context for graph-based financial crime detection. We introduce Temporal-Context Edge Enrichment (TCEE), which uses Chronos-2, a pretrained TSFM, to encode account transaction histories and enables each transaction to condition the retrieval of account-history context on its own features. Unlike approaches that assign all transactions of an account the same representation, TCEE constructs a transaction-specific representation conditioned on the transaction itself. These edge representations can be classified directly or given to a graph neural network (GNN) for additional structural information. Experiments on six AMLWorld anti-money laundering datasets show that TCEE without a GNN backbone improves PR-AUC over the strongest evaluated baseline on five datasets by 10.8% on average, while reaching its best validation PR-AUC 6.0 times faster end-to-end on average across all six datasets, including the one-time cost of Chronos-2 encoding. Adding a GNN backbone improves PR-AUC by a further 4.8% over TCEE without the backbone on the four datasets trained to convergence, and 2.4% on average across all six. On a real Ethereum phishing graph, Chronos-PNA improves PR-AUC by 14.5% over the strongest evaluated baseline, while full TCEE-PNA improves it by 17.8%, showing that Chronos-2 representations can also transfer beyond anti-money laundering when combined with graph learning.