Unsupervised Time Series Anomaly Prediction with Importance-based Generative Contrastive Learning

TL;DR

The paper tackles unsupervised time series anomaly prediction by addressing the lack of labeled anomaly precursors and the explosion of potential precursor patterns. It introduces IGCL, combining diffusion-based anomaly precursor generation, an overlapping window-based temporal convolutional network for contextual representations, a memory bank of important negative samples, and a contrastive learning objective guided by theoretical MMD/Markov insights. Key contributions include a formal problem formulation with theoretical results, a scalable memory-enabled precursor generation mechanism, and strong empirical performance across nine real-world datasets. The work enables robust early warning for safety-critical systems without labeled data and offers a scalable framework for multi-variable time series analysis.

Abstract

Time series anomaly prediction plays an essential role in many real-world scenarios, such as environmental prevention and prompt maintenance of cyber-physical systems. However, existing time series anomaly prediction methods mainly require supervised training with plenty of manually labeled data, which are difficult to obtain in practice. Besides, unseen anomalies can occur during inference, which could differ from the labeled training data and make these models fail to predict such new anomalies. In this paper, we study a novel problem of unsupervised time series anomaly prediction. We provide a theoretical analysis and propose Importance-based Generative Contrastive Learning (IGCL) to address the aforementioned problems. IGCL distinguishes between normal and anomaly precursors, which are generated by our anomaly precursor pattern generation module. To address the efficiency issues caused by the potential complex anomaly precursor combinations, we propose a memory bank with importance-based scores to adaptively store representative anomaly precursors and generate more complicated anomaly precursors. Extensive experiments on seven benchmark datasets show our method outperforms state-of-the-art baselines on unsupervised time series anomaly prediction problems.

Figures (4)

• Figure 1: Our motivation.
• Figure 2: The overall architecture.
• Figure 3: The overlapping window-based temporal convolutional network with a kernel size $k$ of 2.
• Figure 4: IGCL can output larger scores on the precursors that start to deviate from the normal ahead of the severer anomalies.