Research map / Probabilistic and scientific modeling / Causal discovery
A Simple yet Scalable Granger Causal Structural Learning Approach for Topological Event Sequences
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes S2GCSL, a linear-kernel topological Hawkes-style Granger causal structure learner optimized by Adam with sparsity and acyclicity constraints, targeting scalable alarm root-cause analysis in telecom networks.
Problem
Learning Granger causal graphs among alarm event types from topological event sequences in telecommunication networks, where existing methods are either inefficient or do not scale to large numbers of event types and devices.
Approach
Model event activation with a linear kernel over a topological (geodesic-distance) device adjacency inside a Poisson counting model, and optimize the exact log-likelihood by gradient descent (Adam) with L1 sparsity and NOTEARS-style acyclicity penalties, instead of gradient-free or neural point-process optimization.
Key results
- On the metropolitan telecom alarm data, S2GCSL reached F1 0.40±0.06 and SHD 60.6±6.59, versus 0.29 F1 for the next best (CAUSE) and 0.23 for TNPAR.
- S2GCSL ran on 100-event-type synthetic problems in about 9 minutes, while TNPAR, CAUSE, PCMCI, MLE_SGL and ACD failed to finish within one hour at that scale.
- On the real-world data S2GCSL took 737 s versus 4604 s for TNPAR and 7209 s for CAUSE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use S2GCSL instead of TNPAR | Granger causal discovery on metropolitan telecommunication network alarm data · Metropolitan telecommunication network alarm data | F1 Score: 0.23±0.06 -> 0.40±0.06 | Helped |
| use S2GCSL instead of CAUSE | Granger causal discovery on metropolitan telecommunication network alarm data · Metropolitan telecommunication network alarm data | F1 Score: 0.29±0.04 -> 0.40±0.06 | Helped |
| use S2GCSL instead of ADM4 | Granger causal discovery on metropolitan telecommunication network alarm data · Metropolitan telecommunication network alarm data | F1 Score: 0.19±0.03 -> 0.40±0.06 | Helped |
| use S2GCSL instead of PCMCI | Granger causal discovery on metropolitan telecommunication network alarm data · Metropolitan telecommunication network alarm data | SHD: 75.5±4.32 -> 60.6±6.59 | Helped |
| use S2GCSL instead of PCMCI | Granger causal discovery on metropolitan telecommunication network alarm data · Metropolitan telecommunication network alarm data | SID: 367±17.1 -> 397±30.2 | Hurt |
| incorporate topological network information | Granger causal discovery methods S2GCSL, CAUSE, and TNPAR · Metropolitan telecommunication network alarm data | F1 Score | Helped |
| use S2GCSL instead of TNPAR | Granger causal discovery on metropolitan telecommunication network alarm data · Metropolitan telecommunication network alarm data | ET(s): 4604s -> 737s | Helped |
| use S2GCSL instead of CAUSE | Granger causal discovery on metropolitan telecommunication network alarm data · Metropolitan telecommunication network alarm data | ET(s): 7209s -> 737s | Helped |
| use S2GCSL instead of ADM4 | Granger causal discovery on metropolitan telecommunication network alarm data · Metropolitan telecommunication network alarm data | ET(s): 861s -> 737s | Helped |
Surprising findings
- PCMCI achieved the lowest SID-related competitiveness only by producing very sparse graphs, so its SID (367) was lower than S2GCSL's (397) on real data.
Limitations
- Linear kernel may not capture non-linear or non-stationary interactions in dynamic networks.
- Real-world hyperparameters (lambda1, lambda2, pruning threshold) are tuned per setting and depend on sample size, with heuristic selection.
- Evaluated only on one real-world telecom dataset and synthetic data.
Experimental setup
Datasets: synthetic topological event sequences (Erdos-Renyi generated) NeurIPS 2023 competition metropolitan telecommunication network alarm data
Benchmarks: synthetic topological event sequences NeurIPS 2023 telecom alarm dataset (Phase 2 real-world)
Metrics: F1 score SHD SID wall-clock execution time
Baselines: TNPAR ADM4 CAUSE PCMCI MLE_SGL ACD
Builds on: topological Hawkes process NOTEARS ADM4 Adam
Closest papers by meaning
- TNPAR: Topological Neural Poisson Auto-Regressive Model for Learning Granger Causal Structure from Event Sequences (AAAI 2024 · similarity 0.89)
- Causal Discovery from Event Sequences by Local Cause-Effect Attribution (NeurIPS 2024 · similarity 0.87)
- MOCHA: Discovering Multi-Order Dynamic Causal Structure in Temporal Point Processes (NeurIPS 2026 · similarity 0.85)
- GCAD: Anomaly Detection in Multivariate Time Series from the Perspective of Granger Causality (AAAI 2025 · similarity 0.85)
- DCILP: A Distributed Approach for Large-Scale Causal Structure Learning (AAAI 2025 · similarity 0.84)
- Root Cause Analysis in Microservice Using Neural Granger Causal Discovery (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.