CrGSTA: Cross-domain Root causal Graph Spatial-Temporal Attention Network

ICLR 2026 Conference Submission13845 Authors

Keywords: Granger Causality, encoder-decoder, root cause analysis, anomaly detection, attention, fourier

TL;DR: Root cause analysis using neural granger causality through scalable, efficient, graph spatio-temporal cross domain attention encoder-decoder model.

Abstract:

Modern monitoring systems generate massive, high-dimensional time series
where failures rarely remain isolated but cascade across interdependent components.
Identifying their true origins requires more than anomaly detection; it requires
interpretable models that disentangle causal structure from noisy signals.
While Granger causality has gained traction for root cause analysis (RCA), existing
neural methods often rely on multilayer perceptrons applied independently at
each time step, which increases parameter counts, struggles with long-range dependencies,
and overlooks seasonal and periodic patterns. We introduce CrGSTA
(Cross-domain Root causal Graph Spatial-Temporal Attention Network), a scalable
and interpretable framework that unifies time- and frequency-domain representations
through cross-domain attention. CrGSTA employs graph-based spatiotemporal
attention to capture directional dependencies, while frequency-aware
features recover periodic structure. A lightweight self-attention decoder reconstructs
dynamics, ensuring deviations are attributed to true root causes rather than
propagated effects. We conduct experiments along three dimensions: temporal
scalability, spatial scalability, and ablations on domain contributions and fusion
strategies. On both the Lotka–Volterra benchmark and the SWaT industrial
dataset, CrGSTA new state of the art achieving up to 13% Avg@10 improvement
by leveraging wider temporal windows with only 8.5M parameters compared to
(200M+) of other baselines. By explicitly coupling temporal and frequency cues,
CrGSTA balances accuracy, interpretability, and efficiency for RCA in complex
monitoring environments, providing a foundation for more resilient and transparent
analysis in real-world systems. https://github.com/crgsta2025/CrGSTA

Primary Area: causal reasoning

Submission Number: 13845