The Review of Socionetwork Strategies Бизнес и менеджмент
DGAB: Enhancing Credit Card Fraud Detection with Dynamic Graph Attribute Binning Representation
Аннотация
Most fraud detection GNNs treat continuous attributes like transaction amounts as ordinary feature vectors. However, this approach often suppresses critical fraud signals during the neighborhood aggregation process. This limitation is particularly severe for heavy-tailed distributions, in which both an unusually large purchase and a small probing transaction are pulled toward the mode by the overwhelming volume of ordinary ones. To address this limitation, we propose Dynamic Graph Attribute Binning (DGAB). This framework jointly learns adaptive bin boundaries in a log-transformed domain and projects numerical values into bin-specific embedding subspaces before message passing. We couple DGAB with a gated temporal attention backbone that dynamically balances current node evidence with historical neighborhood context. To further exploit unlabeled interaction data, we incorporate a semi-supervised risk-embedding strategy that enriches node representations beyond the labeled training set. Experiments on a real-world financial dataset and two public graph benchmarks show that DGAB consistently outperforms competitive baselines.