VDDFormer: A Variable Dependency Discrepancy-Based Transformer for Multivariate Time Series Anomaly Detection

The dynamics of multivariate time series (MTS) data are jointly characterized by its nonlinear temporal dependencies and complex variable dependencies, making unsupervised time series anomaly detection a challenging task. Existing methods primarily rely on prediction or reconstruction errors, neglecting the valuable information within the variable dependencies. In this paper, we propose a variable dependency discrepancy-based Transformer (VDDFormer) for unsupervised MTS anomaly detection. VDDFormer comprises a variable correlation encoder, a temporal dependency encoder, and a reconstruction decoder. The variable correlation encoder capitalizes on a variable dependency attention mechanism, which employs self-attention to learn the global variable dependencies; meanwhile, the local variable dependencies are captured by the adaptive correlation matrix. The global and local variable dependencies are then used to compute the variable dependency discrepancy as a new intrinsic property to distinguish between normal and abnormal patterns. By integrating this new discrepancy with the reconstruction error, the model effectively enhances its anomaly differentiation capability. Extensive experiments on five real-world anomaly detection datasets demonstrate that VDDFormer effectively and robustly detects group anomaly patterns by leveraging the variable dependency discrepancy and achieves state-of-the-art performance on four out of the five datasets.