An attention-based balanced variational autoencoder method for credit card fraud detection

Credit card fraud detection poses a significant role in both academia and industries. Conventionally, manual checking is time-consuming and not accurate. With the advent of artificial intelligence (AI), numerous machine-learning approaches have emerged. Most of them are proven effective and efficient, especially those based on attention mechanisms. However, the highly imbalanced distribution of normal and fraudulent transactions increases the hardness of further performance improvement. So far, there exist three issues in current contributions: 1) the existing oversampling methods did not fully consider the imbalanced distribution theoretically; 2) the current methods neglected the importance of features and did not mine helpful information from them; 3) the layers of current attention-based methods are too shallow to capture enough characteristics of data. To address the above issues, we propose a novel method: Balanced-Variational AutoEncoder-Attention (Bal-VAE-Attention), which can significantly boost the existing fraud detection results. In our approach, we design a novel oversampling method that considers the imbalanced data distribution. Moreover, we apply an automatic feature selection procedure. Also, we deploy a deep multi-head attention structure to depict the complex inner structure of fraudulent data. We implement abundant experiments on two open-source credit card fraud datasets. Through substantial experiments and ablation study, we prove our proposed method’s effectiveness. To the best of our knowledge, it is superior to other state-of-the-art credit card fraud detection baselines. • Creating a novel loss function that has been proven effective. • Constructing a variational autoencoder method to resolve the data imbalance issue. • Building an attention-based model to improve credit card fraud detection efficacy. • Comparison of the proposed method with state-of-the-art fraud detection baselines.