Sparse Neural Networks for Anomaly Detection in High-Dimensional Time Series

Online anomaly detection in time series is an important component for automated monitoring. In many applications, time series are highdimensional with tens or even hundreds of variables being monitored simultaneously. We note that existing anomaly detection approaches based on recurrent autoencoders may not be very effective for high-dimensional time series. In this work, we propose a simple, yet effective, extension to such approaches for high-dimensional time series. Our approach combines the advantages of non-temporal dimensionality reduction techniques and recurrent autoencoders for time series modeling through an end-to-end learning framework. The recurrent encoder gets sparse access to the input dimensions via a feedforward layer while the recurrent decoder is forced to reconstruct all the input dimensions, thereby leading to better regularization and a robust temporal model. The autoencoder thus trained on normal time series is likely to give a high reconstruction error, and a corresponding high anomaly score, for any anomalous time series pattern. We prove the efficacy of the proposed approach through experiments on a public dataset and two real-world datasets with significant improvement in anomaly detection performance over several baselines. We observe that the proposed approach is able to perform well even without knowledge of relevant dimensions carrying the anomalous signature in a high-dimensional setting.