Fraud-Preserving Dual-Horizon Temporal Heterogeneous Graph Contrastive Learning for Credit Card Fraud Detection

Learning from financial transaction graphs requires both recency sensitivity and retention of sparse, long-range behavioral evidence. We study a fraud-preserving adaptive augmentation (FPAA) mechanism in a dual-horizon heterogeneous graph encoder. Each transaction receives strictly earlier user-, merchant-, and device-linked messages, with a recent exponential-decay path and a long-memory path. Training combines focal classification, TH-GCL-style structurally and temporally perturbed positive pairs, and class-conditioned edge survival applied only to observed training labels. On a fully synthetic, reproducible 30,000-event stream (2.40% fraud; chronological 60/20/20 split), the proposed dual-horizon FPAA model achieves average precision 0.427 ± 0.038 across three training seeds, versus 0.421 ± 0.032 for uniform perturbations and 0.441 ± 0.032 for the dual-horizon supervised encoder. The short-memory-only supervised encoder achieves 0.368, indicating that retaining long-range history is consequential in this benchmark. A logged view-survival audit confirms that FPAA preserves 96.0% of fraud-linked training edges under structural augmentation, compared with 71.2% for uniform dropout. Across three independently generated synthetic streams with the same simulator design, FPAA mean AP is 0.423, against 0.417 for ordinary contrastive augmentation; a causally history-enriched LightGBM reaches 0.502. Hence, protecting fraud edges changes view integrity but provides a modest predictive gain, and historical information alone explains much of the initially observed graph advantage.