A Deep Learning Hybrid RNN-LSTM Model for Credit Card Fraud Detection

To address the limitation that traditional classification models perform poorly when faced with highly imbalanced data in credit card fraud detection, this paper puts forward a joint model that integrates a Recurrent Neural Network - Long Short - Term Memory (RNN - LSTM) architecture with a hybrid sampling strategy. By conducting a systematic analysis of the category distribution characteristics of real transaction datasets and combining with over-sampling or under-sampling data balance processing, a deep neural network capable of learning temporal dependencies is established. Experimental results demonstrate that the optimized RNN - LSTM model achieves breakthrough performance on the publicly available EU credit card transaction dataset. Compared with traditional classifiers such as K - Nearest Neighbors (KNN), AdaBoost, and the standalone LSTM model, the proposed approach improves the F1 - score by 5.31% (reaching 0.8544), maintains an overall accuracy of 99.96%, and achieves a fraud recall rate of 80.56%, effectively balancing the crucial trade - off between false alarm control and positive sample identification. Furthermore, multidimensional performance comparisons show that the proposed approach outperforms the benchmark models in precision (84.12%) and the Area Under the Curve (AUC) value (0.932), validating the synergistic effect of structural optimization and the hybrid sampling strategy. This study provides a solution with both theoretical and practical significance for financial risk prevention and control. Moreover, the proposed methodological framework can be extended to anomaly detection tasks involving other types of imbalanced time - series data.