Temporal-Directional Tree-Walk Graph Neural Network for Illicit Transaction Detection

Illicit transactions in large-scale financial networks manifest as rare, directional temporal motifs, which are critical for accurate detection. However, existing graph-based methods often ignore temporal causality, underutilize directional motifs, or rely on ad hoc weighting, limiting minority-class recognition. This article proposes a temporal-directional tree-walk graph neural network (TDTW-GNN), a causal, motif-aware framework for imbalanced edge classification on continuous-time dynamic graphs (CTDGs). In particular, TDTW-GNN constructs node memories via a temporal graph network (TGN) backbone, performs temporal, direction-aware tree sampling guided by multiorder proximity and Hawkes intensities, encodes hierarchical structure with modified preorder tree traversal (MPTT) through a dual-branch transformer-principal neighborhood aggregation (PNA) network, and fuses in-/out-flow motif and node embeddings for classification, stabilized by a hybrid Dice-focal loss. Experiments on Ethereum phishing, international business machines (IBM) anti-money laundering (AML), and Bitcoin Alpha show substantial minority-class improvements (e.g., + 34.9% $F1$ /+ 40.8% area under the precision-recall curve (PR-AUC) on Ethereum Phishing Detection (ETH); + 10.1% $F1$ /+ 2.5% PR-AUC on AML), and robust performance under sparse labels (ETH 10% labels: $F1$ % and 78.2%, PR-AUC 81.3%). These results, together with ablation studies, demonstrate that TDTW-GNN effectively captures temporal motifs and directional flows, significantly enhancing illicit transaction detection in dynamic financial networks.