A novel approach for online anomaly detection in multivariate time series data that addresses issues of temporal causality and spurious correlations by grouping correlated channels and employing a causal mixer.
Researchers propose a cluster-aware causal mixer designed for online anomaly detection in multivariate time-series data. The method addresses limitations in existing MLP-based models by grouping channels based on their inter-channel correlations into distinct clusters, each receiving a dedicated embedding. A causal mixer is introduced to integrate this information while strictly maintaining temporal causality. Furthermore, the work details a sequential anomaly-scoring method that accumulates evidence over time to refine anomaly boundaries. The model is designed to operate in an online fashion, making it highly suitable for real-time applications. Experimental results show superior performance across six public benchmark datasets.