Time series anomaly detection via temporal relationship graphs and adaptive smoothing
Rongfei Ma, Yuhao Ma, Xiufeng Liu
Abstract
Anomaly detection in time series data is crucial across diverse domains yet challenging due to complex temporal dependencies and high dimensionality . Existing methods often fail to capture a holistic view of these dependencies, overlooking subtle anomalies. This paper introduces a novel framework integrating multifaceted temporal correlation modeling with efficient dimensionality reduction for comprehensive anomaly detection. We construct three distinct Temporal Correlation Graphs (TCGs) – Similarity, Causality, and Synchronization – capturing diverse temporal dependencies beyond pairwise similarity . We seamlessly incorporate Reverse Piecewise Aggregate Approximation (RPAA) within the TCG construction , reducing dimensionality while preserving essential temporal features. Our framework uses a diverse set of statistical, graph-theoretic, and temporal metrics combined with a context-aware scoring system leveraging TCG clusters, enabling accurate detection of both point-based and event-based anomalies. Extensive evaluations on real-world and synthetic datasets demonstrate superior performance, achieving up to a 17% improvement in the F1-score compared to state-of-the-art techniques across a wide range of anomaly types. The statistical significance of these improvements is confirmed through a Wilcoxon signed-rank test.