METrIC-G: Multigraph Enrichment and Time-Aware Message Passing for Imbalanced Classification in Transaction Graphs

Abstract

Modern digital ecosystems comprise entities connected through multiple directed interactions, forming complex transaction graphs. These graphs pose significant challenges for node and edge classification due to domain-specific interaction patterns spanning multiple hops and related transactions, multigraph structures with rich edge features, and often severe class imbalance. State-of-the-art graph neural networks often fail to fully exploit multigraph structures and effectively capture such structural-temporal patterns, resulting in suboptimal classification performance. To address these limitations, we introduce METrIC-G, a novel approach designed to capture multi-hop transaction patterns among connected entities occurring in close temporal proximity within a graph-neural-network-based architecture. METrIC-G builds expressive time-aware node- and edge-features from transaction sequences and time-based statistics. METrIC-G effectively identifies symbolic, class-discriminative patterns among transaction outliers to mitigate class imbalance. METrIC-G employs a time-aware message-passing mechanism that adaptively weights neighbor contributions based on a learnable temporal decay function. We experimentally demonstrate that METrIC-G outperforms all baselines by at least 5.15 percentage points (p.p.) in the minority class F1-score and at least 2.92 p.p. in macro F1-score across all datasets.