Fraud detection on the time series data with machinelearning and deep learning techniques

Credit cards are becoming more popular due to the increased use of technology and a shift toward e-commerce. The development of fraud detection methods has reduced the ratio of credit card frauds, and the total amount of money lost to fraudulent transactions increases as the user base grows. Furthermore, newer financial fraud methods are introduced daily, making it a dynamic and evolving problem that requires the research community's attention through pattern recognition. This study presents a credit card fraud detection approach that addresses several challenges in designing a fraud detection algorithm. Because fraudulent transactions are far fewer in number than regular transactions, the class imbalance is a significant issue. An approach incorporating the synthetic oversampling of the minority class is chosen to obtain a reasonable sample size. The main contribution here is a hybrid feature selection approach, which combines the outputs of the filter based and wrapper-based methods to get optimal features. This proposed technique was compared to three other feature selection techniques; outperformed all those techniques on the four machine learning algorithms, and then applied different balancing techniques to increase the performance. A deep learning model (LSTM) was also trained on the dataset, using SMOTE sampling method with minority classes increased, achieving a remarkably high result. The proposed model outperforms its competitors.