High-Performance AI Framework for Real-Time Anomaly Detection in Digital Payment Streams

The rapid growth of digital payments necessitates effective automated fraud detection. This project utilized a large-scale transactional dataset (over 6.3 million records) to develop a model capable of distinguishing between legitimate and fraudulent transactions, a task severely complicated by extreme class imbalance (where fraud constitutes less than 0.13\%of the data). A Logistic Regression model was implemented as the primary Artificial Intelligence (AI) classifier. The model achieved a very high overall Accuracy of 99.93%. However, a detailed analysis of the performance on the minority class revealed a critical limitation: the model attained a low Recall score of 66.57%, resulting in 825 False Negatives (fraudulent transactions missed) in the test set. This finding concludes that while the model efficiently handles most legitimate transactions, its linear nature and inability to adequately address class imbalance render it insufficient for robust, real-world deployment in financial security. The project successfully established a foundational AI benchmark, clearly quantifying the challenge of achieving reliable fraud capture in highly imbalanced payment environments.