Hyperbolic adaptive spatial-aware multivariate time series anomaly detection

Abstract Existing multivariate anomaly detection methods suffer from practical industrial limitations stemming from distribution assumptions, volatile data and scarce reliable anomaly labels. This study proposes a hyperbolic adaptive spatial-aware multivariate anomaly detection method aimed at enhancing accuracy, robustness and interpretability in real-world applications. It first constructs an adaptive hyperbolic pre-training framework to embed coupled spatio-temporal hierarchical time-series features and mine latent cross-variable correlations. Next, an adaptive graph structure discovery and dynamic sparsification module removes rigid topology restraints, autonomously learning inherent data structures and pinpointing the spatio-temporal locations of anomalies for causal interpretation. A multi-level attention module and small-sample dynamic threshold algorithm further improve model stability when handling complex signals. Built upon self-adaptive graph topology, the unified end-to-end framework integrates anomaly recognition, diagnosis and interpretation with dedicated discrimination and correction mechanisms. Experiments demonstrate 10%–30% performance improvements across diverse detection benchmarks, verifying the method’s effectiveness and advancing interpretable AI research for anomaly detection.