Credit Card Fraud Detection Using Bidirectional LSTM with Attention Mechanism on Sequential Spending Behavior

Credit card fraud detection is structurally a sequence-modelling problem masquerading as a static classification problem: whether a given transaction is fraudulent depends substantially on its position within a cardholder's recent spending pattern-an otherwise unremarkable purchase becomes suspicious chiefly because of what immediately preceded it-information a feature-engineered snapshot model discards by construction. Bidirectional Long Short-Term Memory (BiLSTM) networks process each cardholder's transaction sequence in both temporal directions, and an attention mechanism layered atop the BiLSTM's hidden states learns to weight which positions in that sequence are most informative for the current transaction's fraud probability, rather than treating all sequence positions as equally relevant. This paper applies a BiLSTM-Attention model (128 hidden units per direction, additive Bahdanau-style attention, 20-transaction sliding window) to 8,400,000 transactions from 420,000 cardholders (mean 20 transactions per cardholder), of which 15,120 are confirmed fraudulent (0.18% base rate), benchmarking against logistic regression, Random Forest, and a standard unidirectional LSTM without attention. On a held-out test set, BiLSTM-Attention achieves AUC-PR = 0.864 and 84.4% recall at a fixed 90% precision operating point, exceeding the standard LSTM (AUC-PR 0.764; 13.1% relative improvement), Random Forest (0.684; 26.3%), and logistic regression (0.524; 64.9%). Attention-weight analysis finds the model assigns 38.4% of total attention mass to the three most recent transactions in each 20-transaction window-2.56 times the 15% a position-blind uniform allocation would assign-confirming the model has learned a recency-weighted decision pattern consistent with how fraud investigators describe genuine suspicious-sequence detection, rather than relying on transaction-level features alone. Deployed for real-time authorization scoring at a mean inference latency of 42 milliseconds, the system reduced realised chargeback losses from 4,840,000 to 2,120,000 over a comparable evaluation period (−56.2%) while simultaneously reducing the false-decline rate (legitimate transactions incorrectly blocked) from 2.84% to 1.12% (−60.6%), improvements achieved jointly rather than traded against each other. Cold-start performance for cardholders with fewer than five historical transactions (AUC-PR 0.524, a 39.4% relative degradation), six-month post-deployment concept drift absent retraining (AUC-PR falling to 0.724, -16.2%), reduced cross-border-transaction performance (AUC-PR 0.684, -20.8%), and only 50.0% detection of deliberately structured sub-threshold transaction sequences each bound the model's current scope and motivate specific, named operational mitigations described in Section VII.

Hali tarjima qilinmagan