於三個公開資料集（S-FFSD、YelpChi、Amazon）以三個隨機種子評估。多尺度方法在詐欺群聚、圖稀疏且標註部分缺失時效益最顯著：S-FFSD 測試 AUC 達 0.986，最差種子仍達 0.984，遠較單尺度骨幹（0.892 至 0.990）穩定；交叉驗證估計（約 0.81）較低，存在泛化落差，本文全程均一併報告。本研究不宣稱全面超越先前方法：YelpChi 上有高階方法表現更佳，Amazon 上單尺度即具競爭力。;Financial fraud detection poses three structural difficulties: fraud is rare， adversarial behaviour drifts over time， and trustworthy labels are scarce and costly. A practical detector must exploit the relational structure between transactions， learn from few labels， and adapt to drift without forgetting prior knowledge.
This thesis extends the Risk-aware Gated Temporal Attention Network (RGTAN) along three axes. First， a multi-scale transformer graph-attention head (MS-GTGA) runs attention at three neighbourhood resolutions -- all edges， the top-50%， and the top-25% strongest -- and fuses them with a learnable gate. Second， the model is trained in a staged Active-Learning to Continual-Learning framework: an actively-trained teacher warm-starts a student regularized by Elastic Weight Consolidation， Knowledge Distillation， and Experience Replay. Third， the conditions under which each component helps are characterised empirically.
Evaluation on three public benchmarks (S-FFSD， YelpChi， Amazon) with 3-seed mean plus or minus standard deviation shows the multi-scale model is strongest where fraud is clustered and the graph is sparse. On S-FFSD the combined model achieves test AUC 0.986 plus or minus 0.002 and never falls below 0.984 across seeds; the single-scale baseline ranges from 0.892 to 0.990， so the advantage is largest in the worst case. A measured concentration of fraud-fraud edges in the pruned sub-graph provides mechanistic evidence that pruning learns to denoise; a same-parameter [1，1，1] control narrows the accuracy claim to the architecture rather than pruning alone. Continual learning reduces forgetting under a chronological split; under i.i.d. cross-validation forgetting is near-zero and serves as a sanity check only. A cross-validation estimate (oof_ap approximately 0.81) versus final test AP (0.986) generalization gap is reported throughout.
The contributions are (1) the MS-GTGA multi-scale gated-fusion head， (2) the staged AL-to-CL training framework， and (3) an empirical characterisation of when multi-scale aggregation helps. The backbone is inherited from RGTAN; a high-order method exceeds ours on YelpChi and single-scale RGTAN is competitive on Amazon.
