Fraud-Preserving Dual-Horizon Temporal Heterogeneous Graph Contrastive Learning for Credit Card Fraud Detection (Academic Journal of Finance and Accounting)

Summary

• On a 30,000-event synthetic stream with a 2.40% overall fraud rate, the proposed dual-horizon FPAA model achieves an average precision of 0.427 ± 0.038 across three training seeds.p.1
• An edge-survival audit confirms that FPAA retains 96.0% of fraud-linked training edges under structural augmentation, compared to 71.2% for uniform edge dropout.p.1
• Adding a long-memory path to the recent exponential-decay path increases supervised average precision from 0.368 to 0.441, representing a 7.29 percentage point improvement.p.8
• At a fixed top-50 queue inspection budget, the dual supervised model recovers 47 fraud cases (94% precision), while FPAA recovers 45 (90% precision).p.10
• When provided with identical causal history features across three independent synthetic event streams, a LightGBM baseline achieves a mean average precision of 0.502 compared to 0.423 for FPAA.p.11

Key insights

1. Retaining long-range historical transaction evidence is critical, as a short-memory-only supervised model achieves an AP of 0.368 compared with 0.441 for the dual-horizon supervised encoder.p.1
2. Class-conditioned structural edge protection preserves 96.0% of fraud training edges but delivers only a modest predictive gain over uniform contrastive perturbations.p.1
3. Standard tabular models without historical context perform poorly, with standalone LightGBM achieving an AP of 0.225 and a tabular MLP reaching 0.244.p.7
4. Exponential time-decay weighting alone can discount valuable historical context, which is counteracted by integrating an unweighted long-memory aggregation path.p.8
5. Evaluating performance under fixed queue inspection budgets reveals that whole-curve AP gains do not always translate to superior top-of-queue precision.p.10
6. Equipping decision-tree baselines with explicit causal historical features yields a mean AP of 0.502, demonstrating that tabular models can effectively exploit structured relational context.p.11
7. The fraud-preserving view modification targets training-time optimization and does not depend on positive label availability at inference or authorization time.p.12