Combining Autoencoders and Deep Learning for Effective Fraud Detection in Credit Card Transactions

The advancement of technologies and the proliferation of new payment services have improved our lives, offering limitless opportunities for individuals and companies in each country to develop their businesses through credit card transactions as a payment method. Consequently, continuous improvement is crucial for these systems, particularly in the classification of fraud transactions. Numerous studies are required in the realm of automated and real-time fraud detection. Due to their advantageous properties, recent studies have utilized different deep learning architectures to create well-fitting models to identify fraudulent transactions. Our proposed solution aims to exploit the robust capabilities of deep learning approaches to identify abnormal transactions. The solution can be presented as follows: To address the imbalanced data set issue, we applied an autoencoder combined with the support vector machine model (ASVM). For the classification phase, we utilize an attention-long short-term memory neural network as a weak learner for the gradient boosting algorithm (GB_ALSTM), comparing it with various techniques, including artificial neural networks (ANNs), convolutional neural networks (CNNs), long short-term memory neural networks (LSTMs), attention-long short-term memory neural networks (ALSTMs), and bidirectional long short-term memory neural networks (BLSTMs). We conducted several experiments on a real-world dataset, revealing promising results in detecting abnormal transactions and highlighting the dominance of our suggested solution over competing models.