Hybrid big data stream processing model for effective fraud detection in cash electronic payment cards

Fraud detection in banking systems is of critical importance due to its vital role in identifying unauthorized transactions and preventing financial losses. The complexity of data structures, high transaction volume, and continuous evolution of fraud patterns pose significant challenges. This paper presents an innovative system-level hybrid framework for fraud detection in electronic payment cards, which integrates unsupervised and supervised algorithms within a novel two-stage streaming pipeline featuring adaptive per-cardholder profiling, Dempster–Shafer-based score fusion, and drift-aware dynamic thresholds. While building upon established algorithms, the primary novelty lies in the architectural design and synergistic integration rather than new base methods. The model first constructs distinct user behavior profiles (based on transaction timing, amounts, and other features) with adaptive thresholds, periodically updated in the batch learning layer to handle concept drift in streaming environments. It then applies a two-stage filtering process: the fast preliminary filter uses unsupervised algorithms K-means clustering and Hidden Markov model (HMM) to rapidly detect anomalies and reduce candidates, selected for their efficiency, scalability, and complementary capture of spatial and temporal deviations. The explicit filter employs supervised algorithms K-nearest neighbors (KNN), Logistic Regression, and Decision Tree to classify remaining suspicious transactions with high precision, chosen for their effectiveness in imbalanced data, interpretability, local pattern recognition, and non-linear modeling. The framework also leverages the maximum difference between consecutive observations (defined as the largest absolute change in normalized scores or features between successive steps, motivated by its sensitivity to abrupt anomalies) as a key discriminative feature. The model is evaluated using classification metrics (cross-validation accuracy, recall, precision, F1-score) for categorical fraud detection and regression metrics (RMSE, MBE) for numerical anomaly prediction. Results on a large credit card transaction dataset show the proposed framework outperforms several state-of-the-art baselines (Random forest, XGBoost, Isolation forest, Autoencoder) with statistically significant improvements ( p < 0.05 or better) in F1-score and recall, while achieving superior detection performance and computational efficiency. By combining multi-algorithm strengths with in-memory stream processing, this research delivers an effective, scalable solution for real-time fraud detection in electronic payment systems with built-in adaptability to evolving patterns.