Angel or devil: Discriminating hard samples and anomaly contaminations for unsupervised time series anomaly detection.

Abstract

Training for unsupervised time series anomaly detection is constantly plagued by the discrimination between harmful anomaly contaminations and beneficial hard normal samples. These two types of samples display similar loss behavior, which conventional loss-based methods struggle to differentiate. To address this issue, we introduce a novel metric that augments traditional loss behavior with parameter behavior, thereby enabling a more granular delineation of anomalous patterns. Parameter behavior is formalized by quantifying the parametric response to minor perturbations in data samples. By exploiting the complementary nature of parameter and loss behaviors, we further introduce PLDA, a dual Parameter-Loss Data Augmentation method. During the training phase of anomaly detection, PLDA dynamically augments the training set via an iterative procedure. It concurrently mitigates anomaly contaminations while amplifying informative hard normal samples. PLDA exhibits an impressive adaptability, enabling it to serve as an additional component that seamlessly integrates with existing anomaly detectors to enhance their performance. Extensive experiments on ten datasets demonstrate that PLDA significantly enhances the performance of four different detectors by up to 8%, outperforming three data augmentation competitors.¹.