STRIC: Stacked Residuals of Interpretable Components for Time Series Anomaly Detection
We present a residual-style architecture for interpretable forecasting and anomaly detection in multivariate time series. Our architecture is composed of stacked residual blocks designed to separate components of the signal such as trends, seasonality, and linear dynamics. These are followed by a Temporal Convolutional Network (TCN) that can freely model the remaining components and can aggregate global statistics from different time series as context for the local predictions of each time series. The architecture can be trained end-to-end and automatically adapts to the time scale of the signals. After modeling the signals, we use an anomaly detection system based on the classic CUMSUM algorithm and a variational approximation of the $f$-divergence to detect both isolated point anomalies and change-points in statistics of the signals. Our method outperforms state-of-the-art robust statistical methods on typical time series benchmarks where deep networks usually underperform. To further illustrate the general applicability of our method, we show that it can be successfully employed on complex data such as text embeddings of newspaper articles.
PDF Abstract
NAB
