An Encode-then-Decompose Approach to Unsupervised Time Series Anomaly Detection on Contaminated Training Data--Extended Version

TL;DR

This work tackles unsupervised time series anomaly detection under contaminated training data by introducing the encode-then-decompose (EDAD) framework, which splits latent representations into stable and auxiliary components. By replacing reconstruction-based objectives with a mutual information (MI) based criterion and employing an asymmetric training and scoring pipeline, EDAD achieves robustness to anomalies in training data. The model uses a self-attentive encoder, a two-branch decomposition with shuffle-based cues, and a teacher–student regularization scheme, optimized via a weighted sum of MI- and reconstruction-inspired losses. Empirical results on eight real-world datasets show competitive to state-of-the-art performance with strong robustness to varying contamination levels and favorable runtime/memory characteristics, underscoring its practical applicability for scalable anomaly detection in diverse domains.

Abstract

Time series anomaly detection is important in modern large-scale systems and is applied in a variety of domains to analyze and monitor the operation of diverse systems. Unsupervised approaches have received widespread interest, as they do not require anomaly labels during training, thus avoiding potentially high costs and having wider applications. Among these, autoencoders have received extensive attention. They use reconstruction errors from compressed representations to define anomaly scores. However, representations learned by autoencoders are sensitive to anomalies in training time series, causing reduced accuracy. We propose a novel encode-then-decompose paradigm, where we decompose the encoded representation into stable and auxiliary representations, thereby enhancing the robustness when training with contaminated time series. In addition, we propose a novel mutual information based metric to replace the reconstruction errors for identifying anomalies. Our proposal demonstrates competitive or state-of-the-art performance on eight commonly used multi- and univariate time series benchmarks and exhibits robustness to time series with different contamination ratios.

Figures (10)

• Figure 1: Autoencoders (AE) vs. Encode-then-Decompose Anomaly Detection (EDAD).
• Figure 2: Framework pipeline. The data preprocessing component is shared by the offline training and online detection stages.
• Figure 3: EDAD overview.
• Figure 4: Regularization.
• Figure 5: Effect of contamination ratio (CR).