Multi-scale temporal modeling for financial fraud detection using variational autoencoders

The explosive growth of digital payment systems has accelerated the adoption of machine learning (ML) in financial anomaly detection. Existing methods attempt to learn the characteristics of anomalous samples from large-scale datasets; however, due to the scarcity of fraudulent transactions, these methods often suffer from limited generalization. In this paper, we present MS-VAE, a novel framework for detecting anomalous financial transactions via a Multi-Scale Variational AutoEncoder. MS-VAE leverages multi-scale temporal convolution networks to extract hierarchical features at different granularities and constructs a variational latent space with Gaussian mixture priors to model the multimodal characteristics of normal transaction behavior. It employs adaptive weighted fusion mechanisms to automatically adjust the contributions of different temporal scales, incorporates self-attention mechanisms to capture long-range dependencies, and introduces temporal consistency constraints to ensure smooth latent representations. As a result, MS-VAE can effectively distinguish anomalous patterns that deviate from the learned distributions of normal behavior. Extensive experiments on three real-world financial datasets validate the effectiveness of MS-VAE, demonstrating an average F1-score improvement of 8.3% over state-of-the-art baselines.