Time Series Causal Discovery via Instance-specific Modeling and Intervention-based Pretraining

Biao Ouyang, Tengxue Zhang, zhihao zhuang, Yang Shu, Chenjuan Guo, Bin Yang

Keywords: Time Series Causal Discovery, Root Cause Identification

Abstract: Causal discovery in time series is crucial for downstream tasks such as tracing the root causes of anomalies. Structural Causal Models (SCMs) provide a principled way to formalize the generative processes of observed data. However, learning generalized causal models for time series remains challenging due to the lack of detailed modeling and limited generalization ability. In this paper, we propose \textbf{T-Caus}, a novel pretraining framework designed to improve the generalization of \textbf{T}ime series \textbf{Caus}al models across diverse downstream tasks. To capture complex temporal causal dependencies, T-Caus introduces a hierarchical instance-specific temporal causal discovery framework that employs a dual-scale iterative attention to enhance window-level causal relationships, and a Gaussian mixture with an instance-level routing mechanism to handle heterogeneous exogenous distributions. To further address distribution shifts across time series, T-Caus adopts generalizable causal learning with causal invariance, which explicitly leverages intervention-based learning and a causal mixup strategy to promote stable causal discovery and stronger generalization. Extensive experiments on multiple real-world out-of-distribution (OOD) datasets demonstrate that T-Caus exhibits strong generalization, achieving superior performance in both causal discovery and root cause identification. The code and datasets are available at the \href{https://anonymous.4open.science/r/T-Caus-B0CD}{link}.

Primary Area: learning on time series and dynamical systems

Submission Number: 9195