Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RealTCD: Temporal Causal Discovery from Interventional Data with Large Language Model

Peiwen Li, Xin Wang, Zeyang Zhang, Yuan Meng, Fang Shen, Yue Li, Jialong Wang, Yang Li, Wenweu Zhu

TL;DR

This work tackles temporal causal discovery in industrial settings where interventional targets are unavailable and much of the system knowledge is encoded in text. It introduces RealTCD, a two-module framework combining a score-based temporal causal discovery method with an LLM-guided meta-initialization to inject domain knowledge from textual data. The optimization uses an augmented Lagrangian with learnable masks to handle unknown interventions and enforce acyclicity, enabling joint learning of the temporal graph and intervention targets. Across synthetic SVAR-like data and a real data-center dataset, RealTCD outperforms strong baselines on both structural metrics and domain-relevant causal relations, demonstrating practical potential for root-cause analysis, anomaly detection, and IT operations optimization.

Abstract

In the field of Artificial Intelligence for Information Technology Operations, causal discovery is pivotal for operation and maintenance of graph construction, facilitating downstream industrial tasks such as root cause analysis. Temporal causal discovery, as an emerging method, aims to identify temporal causal relationships between variables directly from observations by utilizing interventional data. However, existing methods mainly focus on synthetic datasets with heavy reliance on intervention targets and ignore the textual information hidden in real-world systems, failing to conduct causal discovery for real industrial scenarios. To tackle this problem, in this paper we propose to investigate temporal causal discovery in industrial scenarios, which faces two critical challenges: 1) how to discover causal relationships without the interventional targets that are costly to obtain in practice, and 2) how to discover causal relations via leveraging the textual information in systems which can be complex yet abundant in industrial contexts. To address these challenges, we propose the RealTCD framework, which is able to leverage domain knowledge to discover temporal causal relationships without interventional targets. Specifically, we first develop a score-based temporal causal discovery method capable of discovering causal relations for root cause analysis without relying on interventional targets through strategic masking and regularization. Furthermore, by employing Large Language Models (LLMs) to handle texts and integrate domain knowledge, we introduce LLM-guided meta-initialization to extract the meta-knowledge from textual information hidden in systems to boost the quality of discovery. We conduct extensive experiments on simulation and real-world datasets to show the superiority of our proposed RealTCD framework over existing baselines in discovering temporal causal structures.

RealTCD: Temporal Causal Discovery from Interventional Data with Large Language Model

TL;DR

This work tackles temporal causal discovery in industrial settings where interventional targets are unavailable and much of the system knowledge is encoded in text. It introduces RealTCD, a two-module framework combining a score-based temporal causal discovery method with an LLM-guided meta-initialization to inject domain knowledge from textual data. The optimization uses an augmented Lagrangian with learnable masks to handle unknown interventions and enforce acyclicity, enabling joint learning of the temporal graph and intervention targets. Across synthetic SVAR-like data and a real data-center dataset, RealTCD outperforms strong baselines on both structural metrics and domain-relevant causal relations, demonstrating practical potential for root-cause analysis, anomaly detection, and IT operations optimization.

Abstract

In the field of Artificial Intelligence for Information Technology Operations, causal discovery is pivotal for operation and maintenance of graph construction, facilitating downstream industrial tasks such as root cause analysis. Temporal causal discovery, as an emerging method, aims to identify temporal causal relationships between variables directly from observations by utilizing interventional data. However, existing methods mainly focus on synthetic datasets with heavy reliance on intervention targets and ignore the textual information hidden in real-world systems, failing to conduct causal discovery for real industrial scenarios. To tackle this problem, in this paper we propose to investigate temporal causal discovery in industrial scenarios, which faces two critical challenges: 1) how to discover causal relationships without the interventional targets that are costly to obtain in practice, and 2) how to discover causal relations via leveraging the textual information in systems which can be complex yet abundant in industrial contexts. To address these challenges, we propose the RealTCD framework, which is able to leverage domain knowledge to discover temporal causal relationships without interventional targets. Specifically, we first develop a score-based temporal causal discovery method capable of discovering causal relations for root cause analysis without relying on interventional targets through strategic masking and regularization. Furthermore, by employing Large Language Models (LLMs) to handle texts and integrate domain knowledge, we introduce LLM-guided meta-initialization to extract the meta-knowledge from textual information hidden in systems to boost the quality of discovery. We conduct extensive experiments on simulation and real-world datasets to show the superiority of our proposed RealTCD framework over existing baselines in discovering temporal causal structures.
Paper Structure (47 sections, 1 theorem, 7 equations, 3 figures, 6 tables, 1 algorithm)

This paper contains 47 sections, 1 theorem, 7 equations, 3 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Dynamic Causal Graphical Model
  4. Intervention
  5. Method
  6. Score-based Temporal Causal Discovery
  7. Raw data
  8. Model conditional densities
  9. Score for unknown interventional targets
  10. Maximize the score
  11. LLM-guided Meta Initialization
  12. Prompts of LLMs
  13. Guide temporal causal discovery from data
  14. Extract weight
  15. Experiments
  16. ...and 32 more sections

Key Result

theorem 1

Suppose $\mathcal{I}^*$ is such that $\mathcal{I}_1^*:= \emptyset$. Let $\mathcal{G}^*$ be the ground truth temporal DAG and $(\hat{\mathcal{G}},\hat{\mathcal{I}})\in arg max_{\mathcal{G}\in DAG,\mathcal{I}}\mathcal{S}(\mathcal{G},\mathcal{I})$. Under the assumptions that: 1) the density model has e

Figures (3)

  • Figure 1: The framework of our proposed method RealTCD. Given the system textual information and temporal data without interventional targets, the LLM-guided Meta-Initialization module leverages LLMs to extract the domain knowledge and obtain the potential causal relationships as the initialization adjacency matrix $M_0^\mathcal{G}$. Then, the Score-based Temporal Causal Discovery module utilizes an augmented Lagrangian process to optimize the score for unknown interventional targets under constraints, where the $\Lambda_0$ is initialized with $M_0^\mathcal{G}$. In this way, the proposed RealTCD leverages the system textual information to discover temporal causal relationships without interventional targets.
  • Figure 2: A typical data center cooling system diagram.
  • Figure 3: Showcases of the results on synthetic data.

Theorems & Definitions (1)

  • theorem 1: Unknown targets temporal causal DAG identification