PASTA: Neural Architecture Search for Anomaly Detection in Multivariate Time Series

Time-series anomaly detection uncovers rare errors or intriguing events of interest that significantly deviate from normal patterns. In order to precisely detect anomalies, a detector needs to capture intricate underlying temporal dynamics of a time series, often in multiple scales. Thus, a fixed-designed neural network may not be optimal for capturing such complex dynamics as different time-series data require different learning processes to reflect their unique characteristics. This paper proposes a Prediction-based neural Architecture Search for Time series Anomaly detection framework, dubbed PASTA. Unlike previous work, besides searching for a connection between operations, we design a novel search space to search for optimal connections in the temporal dimension among recurrent cells within/between each layer, i.e., temporal connectivity, and encode them via multi-level configuration encoding networks. Experimental results from both real-world and synthetic benchmarks show that the discovered architectures by PASTA outperform the second-best state-of-the-art baseline by around 13.6% in the enhanced time-series aware $F_{1}$ score on average, confirming that the design of temporal connectivity is critical for time-series anomaly detection.