Rethinking Heavy Models in Multivariate Time Series Anomaly Detection

ICLR 2026 Conference Submission17464 Authors

Keywords: Multivariate Time Series Anomaly Detection, Effectiveness-Efficiency Trade-off, Resource-Constrained Environment, AUROC, FLOPs

Abstract: Multivariate time series anomaly detection (MTS-AD) is widely used, but real-world deployments often face tight computational budgets that limit the practicality of deep learning. We revisit whether heavy deep models (high-FLOPs architectures) are necessary to achieve strong detection performance in such settings. We conduct a systematic, compute-aware comparison of statistical, classical machine learning, and deep learning methods across diverse MTS-AD benchmarks, measuring detection with AUROC (threshold-free, thus application-agnostic) and cost with FLOPs (a hardware-agnostic proxy enabling fair cross-method comparison). We find that traditional approaches often match or surpass deep models, which appear less frequently among the top performers, and that the effectiveness-efficiency trade-off commonly favors non-deep alternatives under limited budgets. These results indicate that deep learning is not uniformly superior for MTS-AD and that heavy architectures can be counterproductive in resource-constrained deployments. These findings offer practical guidance for practitioners designing anomaly monitoring systems under compute constraints, highlighting cases where lightweight models are sufficient and heavy deep models may be worth the cost.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 17464