Resilient Concurrent Causal Discovery for Topological Event Sequences

Jiyu Tian; Junhao Dong; Mingchu Li; Lingling Fang; Liming Chen; Andreas Holzinger; Zheng Yan; Yew Soon Ong

cs.LGAug 22, 2026

Resilient Concurrent Causal Discovery for Topological Event Sequences

Authors: Jiyu Tian, Junhao Dong, Mingchu Li, Lingling Fang, Liming Chen, Andreas Holzinger, Zheng Yan, Yew Soon Ong

Organizations: School of Software Technology, Dalian University of Technology, Dalian, China · College of Computing & Data Science, Nanyang Technological University, Singapore · School of Computer and Information Engineering, Jiangxi Normal University, Nanchang, Jiangxi, China · School of Computer Science and Artificial Intelligence, Liaoning Normal University, Dalian, China · School of Computer Science and Technology, Dalian University of Technology, Dalian, China · Human-Centered AI Lab, FTEC, Department of Ecosystem Management, Climate and Biodiversity, University of Natural Resources and Life Sciences (BOKU), Vienna, Austria · State Key Lab of ISN, School of Cyber Engineering, Hangzhou Institute of Technology, Xidian University, Xi’an, Shaanxi, China · CFAR, IHPC, A*STAR, Singapore

Abstract

Causal discovery on topological event sequences is crucial for ensuring the reliability of networks. However, existing methods struggle to capture the complex causal relationships arising from concurrent events and lack robustness to incomplete event sequences. To address these issues, we propose a resilient concurrent causal discovery method, termed RCCD, enabling robust learning of causal graphs from topological event sequences. Specifically, we first introduce an influence-aware hyperedge causal attention mechanism, which incorporates event duration into the embedding representation, aggregates concurrent event features via hyperedge causal convolution, and injects network prior knowledge to capture the complex many-to-one causal interactions. Furthermore, we design a masked-based alternating causal optimization framework, which forces the model to recover masked event types based on context through self-supervised mask reconstruction, thereby enhancing the resilience of the predictor to missing data. To validate the effectiveness of our method, we conduct extensive experiments on both simulated and real-world telecommunication network datasets. Experimental results demonstrate that the proposed method significantly outperforms existing state-of-the-art methods in both accuracy and robustness, making it more suitable for real-world telecommunication network environments.

Figures & tables

Fig. 1: An illustrative example of topological event sequence causal discovery. Causal discovery for telecom network event sequences can identify causal relationships from discrete alarms generated by devices. Each alarm event includes an event type, a device type, start timestamps, and end timestamps. Network topology describes the physical connections among devices. Concurrent events refer to multiple alarms occurring simultaneously within the same time window, which may collaboratively produce complex causality. Incomplete observations mean that some alarms are not recorded due to network instability. Causal graph is a directed acyclic graph where nodes represent event types and directed edges indicate that one type directly triggers another type.

Fig. 2: Overview of the RCCD approach, which consists of two main components: Influence-aware Hyperedge Causal Attention (IHCA), which captures complex causal relationships by integrating event duration and concurrent effects while imposing topological constraints; Mask-based Alternation Causal Optimization (MACO), which enhances robustness to missing data through self-supervised reconstruction and optimizes the causal graph with sparsity and acyclicity regularization, ultimately generating an accurate directed acyclic causal graph.

Methods	15V-30N-Synthetic				20V-40N-Synthetic				24V-439N-Microwave				25V-474N-Microwave
Methods	PRE	REC	F1	SHD	PRE	REC	F1	SHD	PRE	REC	F1	SHD	PRE	REC	F1	SHD
PC	0.0141	0.1111	0.0250	77	0.0451	0.3529	0.08	136	0.3243	0.2628	0.2903	142	0.2821	0.223	0.2491	162
GraNDAG	0.1667	0.3333	0.2222	21	0.0500	0.0588	0.0541	34	0.2500	0.0438	0.0745	144	0.4667	0.0473	0.0859	143
CORL	0.1667	0.2222	0.1905	17	0.1429	0.1765	0.1579	32	0.1176	0.0146	0.0260	140	0.2000	0.0338	0.0578	158
THPs	0.4737	1.0000	0.6429	10	0.8000	0.9412	0.8649	5	0.6000	0.0657	0.1184	129	0.6000	0.0811	0.1429	136
TNPAR	0.1000	0.2222	0.1379	25	0.0630	0.4706	0.1111	124	0.2524	0.9781	0.4012	142	0.2462	0.9932	0.3946	153

TABLE I: Performance comparison on Stable datasets

Panel of Figure 4 — Fig. 3: Performance Comparison of Advanced Large Language Models on Synthetic Datasets.

Panel of Figure 5 — Fig. 4: Performance Comparison on Observing Incomplete Sequences.

Panel of Figure 6 — Fig. 5: Sensitivity to Hyperparameters

Scenarios	Componets	15V-30N-Synthetic				20V-40N-Synthetic				24V-439N-Microwave				25V-474N-Microwave
Scenarios	Componets	PRE	REC	F1	SHD	PRE	REC	F1	SHD	PRE	REC	F1	SHD	PRE	REC	F1	SHD
Stable	Baseline	0.5385	0.7778	0.6364	6	0.5417	0.7647	0.6341	15	0.4245	0.5547	0.4810	130	0.4054	0.5067	0.4504	152
Stable	Baseline+IHCA	0.8000	0.8889	0.8421	3	0.9333	0.8235	0.8750	4	0.4130	0.7445	0.5313	135	0.3681	0.7635	0.4967	158
Missing	Baseline	0.3810	0.8889	0.5333	14	0.4211	0.9412	0.5818	23	0.3185	0.7299	0.4435	150	0.3937	0.4256	0.4090	148
	Baseline+IHCA	0.7778	0.7778	0.7778	4	0.8571	0.7059	0.7742	7	0.2807	0.8686	0.4242	156	0.3322	0.6757	0.4454	168
	IHCA+MACO (RCCD)	0.8750	0.7778	0.8235	3	0.8571	0.7059	0.7742	7	0.3449	0.7956	0.4812	149	0.3628	0.7770	0.4946	152

TABLE II: Ablation Study

Causal Discovery

Causal discovery infers cause-and-effect structures from observational and interventional data, distinguishing causal relationships from correlation.

See all 221 papers

TTCD:Transformer Integrated Temporal Causal Discovery from Non-Stationary Time Series Data
Apr 27, 2026Omar Faruque, Sahara Ali, Xue Zheng +1Time-Series Causal Discovery Causal Structure Learning
Department of Information Systems University of Maryland, Baltimore County Baltimore, Maryland, USA · Department of Data Science University of North Texas Denton, Texas, USA · Climate Science Section Lawrence Livermore National Laboratory Livermore, California, USA
Mask2Cause: Causal Discovery via Adjacency Constrained Causal Attention
May 8, 2026Omar Muhammad, Pasupuleti Dhruv Shivkant, Deepak N. SubramaniTime-Series Causal Discovery Causal Structure Learning
1Indian Institute of Science, Bengaluru, India
Causal Discovery on Irregular Time Series
Jul 20, 2026Martim Penim, Ricardo Ribeiro Pereira, Jacopo Bono +3Irregular Time-Series Modeling Time-Series Causal Discovery
1Feedzai, Portugal · 3DCC, Faculdade de Ciêncidas da Universidade do Porto, Portugal · 2Instituto Superior Técnico, Universidade de Lisboa, Portugal +1

Apr 27, 2026J/K move · Enter open · S save

TTCD:Transformer Integrated Temporal Causal Discovery from Non-Stationary Time Series Data

Omar Faruque, Sahara Ali, Xue Zheng, Jianwu Wang

Department of Information Systems University of Maryland, Baltimore County Baltimore, Maryland, USA · Department of Data Science University of North Texas Denton, Texas, USA · Climate Science Section Lawrence Livermore National Laboratory Livermore, California, USA

Time-Series Causal Discovery Causal Structure Learning Non-Stationary Time Series Time Series Transformers

Open paper AskPDF arXiv

The widespread availability of complex time series data in various domains such as environmental science, epidemiology, and economics demands robust causal discovery methods that can identify intricate contemporaneous and lagged relationships in non-stationary, nonlinear, and noisy settings. Existing constraint-based methods often rely heavily on conditional independence tests that degrade for limited data samples and complex distributions, while score-based methods impose strong statistical assumptions. Recent methods address special cases such as change point detection or distribution shifts, but struggle to provide a unified solution. We propose the Transformer Integrated Temporal Causal Discovery (TTCD) Framework, a novel end-to-end approach that learns contemporaneous and lagged causal relations from non-stationary time series. TTCD introduces a Non-Stationary Feature Learner integrating temporal and frequency-domain attention with dynamic non-stationarity profiling, and a custom Causal Structure Learner. A key innovation is reconstruction-guided causal signal distillation, to distill essential causal signals through the reconstruction process of the transformer decoder, which mitigates noise and spurious correlations while preserving meaningful dependencies. The Causal Structure Learner operates on distilled reconstructed signals to infer the underlying causal graph without restrictive assumptions on noise distributions or data generation processes. Experiments on synthetic, benchmark, and real world datasets show that TTCD consistently outperforms state-of-the-art baselines in both accuracy and consistency with domain knowledge, demonstrating the approach's effectiveness for causal discovery in challenging real world contexts.

Abstract

Causal Discovery

Explore similar work

TTCD:Transformer Integrated Temporal Causal Discovery from Non-Stationary Time Series Data

Mask2Cause: Causal Discovery via Adjacency Constrained Causal Attention

Causal Discovery on Irregular Time Series