MSFlow: Fusing Multi-Perspective Scores through Coupling Flows for Time Series Anomaly Detection

Wanghui Qiu, Shiyan Hu, Zhengyu Li, Siyuan Wang, Chenjuan Guo, Bin Yang

Keywords: Time series anomaly detection, ensemble learning, Coupling flows, Progressive supervision, Semi-supervised learning

Abstract: Time series anomaly detection is crucial for ensuring system reliability across various applications ranging from industrial monitoring to financial fraud detection. However, two fundamental challenges remain to be addressed: (1) Model bias caused by the inherent diversity of anomaly patterns; (2) Detection inflexibility caused by the scarcity of anomaly labels. We propose MSFlow (Multi-view Score Fusion via Coupling Flows), which constructs a coupling flow-based ensemble capable of modeling complex joint distributions of multi-dimensional scores through invertible transformations. Leveraging this flexible fusion framework, we strategically select four detection perspectives (clustering and reconstruction in both temporal and frequency domains). The coupling flows learn inter-view dependencies while preserving each perspective's unique detection capabilities, achieving effective integration that simple aggregation fails to accomplish. When labels are available, an uncertainty-guided enhancement mechanism identifies high-disagreement regions in ensemble predictions and selectively refines them through a learned soft router, enabling seamless adaptation from unsupervised to semi-supervised operation. Extensive experiments on 10 univariate and 8 multivariate benchmark datasets demonstrate that MSFlow achieves state-of-the-art performance across diverse anomaly types and label availability scenarios.

Primary Area: learning on time series and dynamical systems

Submission Number: 13134