Time-Series Anomaly Detection Based on Imputation in the Latent Space

Time-series anomaly detection poses a significant challenge, as it requires effectively learning complex normal patterns from the training data to identify unseen anomalies during testing. Recent approaches have concentrated on enhancing self-prediction (such as reconstructing or imputing parts of the input time series) capabilities by capturing intrinsic characteristics such as temporal dependency or inter-correlation in time-series data. The objective is to find a way to measure how well the model predicts, and ideally, this measurement should only show high prediction error during abnormal time intervals, thereby accurately distinguishing them from normal time intervals. How-ever, since conventional approaches make predictions directly in the input space, their inability to pinpoint the exact fluctuation points in the time series can result in high prediction losses, leading to difficulties in anomaly detection. These challenges are exacerbated when the time series exhibits high-frequency traits or when the required prediction target length increases, thereby making the identification of abnormal time intervals even more difficult. To address these issues, we introduce a novel paradigm for time-series anomaly detection by performing imputation at the level of latent representations rather than directly on the input feature space. Our latent imputation module is freed from the necessity of precisely pinpointing fluctuating time steps and can assess the presence of abnormal patterns across consecutive time intervals. Throughout experiments, we demonstrate the superiority of our method over state-of-the-art benchmarks, which is evident in both real-world and synthetic, where multiple types of anomalies are reflected.