Student-t Process Prior Variational Autoencoder for Anomaly Detection in Time Series

Detecting anomalies in time series is vital in areas like web data analysis and fraud detection. While the Variational Autoencoder (VAE) excels at learning non-linear features in time series, it struggles with seasonality due to its assumption of independent latent representations. Additionally, most reconstruction methods need clean data, which isn’t often available. To overcome these issues, we introduce the Student-t process prior variational autoencoders (SPPVAE). This method captures seasonality in long-term time series by integrating time into its kernel function, and its heavy-tailed distribution counters data contamination. Our experiments highlight SPPVAE’s superior performance on multiple datasets against various benchmarks and leading methods.