A hybrid CNN-LSTM-transformer model for window-based anomalous transaction detection in financial data

Detecting fraudulent financial transactions poses significant challenges due to their rarity, adaptability, and high financial impact. Researchers have developed a novel stacked CNN-LSTM-Transformer model for window-based anomaly detection. Tested on the Credit Card Fraud Detection Dataset 2023, which comprises 568,630 anonymized records with 28 PCA-transformed features and transaction amounts, the model was evaluated at varying transaction-level fraud prevalence levels of 10%, 5%, 3%, and 1%.

Utilizing window-level classification with majority voting to handle rare fraud events, the CNN-LSTM-Transformer achieved an impressive 97.07% accuracy, 94.42% precision, 85.70% recall, 89.84% F1-score, 95.32% area under the ROC curve (ROC-AUC), and 90.05% area under the precision-recall curve (PR-AUC) at a 1% transaction-level fraud rate with a 16-window size.

When compared to Logistic Regression, Random Forest, and XGBoost, the CNN-LSTM-Transformer showed no significant difference in F1-score from XGBoost, but outperformed the others in terms of precision and recall. Despite this, XGBoost maintained a slight edge in PR-AUC and precision. The study concludes that while gradient-boosted trees remain effective on PCA-transformed fraud benchmarks, the CNN-LSTM-Transformer is a window-based representation-learning approach better suited for learning transaction patterns rather than modeling chronological behavior.