Research map / Probabilistic and scientific modeling / Causal discovery
TNPAR: Topological Neural Poisson Auto-Regressive Model for Learning Granger Causal Structure from 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 TNPAR, a neural Poisson auto-regressive model with amortized inference that learns Granger causal structure from topological (non-i.i.d.) event sequences by conditioning on a prior network topology.
Problem
Granger causal discovery from event sequences violates the i.i.d. assumption because event sequences on topologically connected network nodes are dependent.
Approach
Model count processes with a neural Poisson auto-regressive generator whose inputs are masked by a causal matrix (over event types and geodesic distances) and a topology matrix, and infer the latent causal matrix via amortized variational inference with Gumbel-Softmax, acyclicity and sparsity regularization.
Key results
- On real alarm data, TNPAR achieves F1 0.54 and recall 0.65, versus 0.44 F1 for THP (the best baseline).
- TNPAR attains the best SID among methods except ACD on real data (241.6 vs. ACD 230.0), while its SHD (105.5) is higher than THP (85.0).
- On simulated data, TNPAR achieves the highest F1 across most parameter settings compared with PCMCI, ACD, MLE_SGL, ADM4, and THP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use topological information across nodes instead of same-node history only (ablation) | TNPAR · Metropolitan Cellular Network Alarm Data | F1: 0.38 -> 0.54 | Helped |
| Model nodes using the topological network rather than merging all nodes into one sequence (ablation) | TNPAR · Metropolitan Cellular Network Alarm Data | F1: 0.33 -> 0.54 | Helped |
| Add acyclicity and sparsity regularization (ablation) | TNPAR · Metropolitan Cellular Network Alarm Data | F1: 0.49 -> 0.54 | Helped |
| Use TNPAR instead of THP | Granger causal discovery · Metropolitan Cellular Network Alarm Data | F1: 0.44 -> 0.54 | Helped |
| Use TNPAR instead of THP | Granger causal discovery · Metropolitan Cellular Network Alarm Data | Precision: 0.58 -> 0.46 | Hurt |
| Use TNPAR instead of THP | Granger causal discovery · Metropolitan Cellular Network Alarm Data | Recall: 0.35 -> 0.65 | Helped |
| Use TNPAR instead of THP | Granger causal discovery · Metropolitan Cellular Network Alarm Data | SID: 266.0 -> 241.6 | Helped |
| Use TNPAR instead of THP | Granger causal discovery · Metropolitan Cellular Network Alarm Data | SHD: 85.0 -> 105.5 | Hurt |
Surprising findings
- On real data TNPAR has a higher SHD than THP and several baselines despite its higher F1, with precision lower than THP, PCMCI and ADM4.
Limitations
- Real-world evaluation uses a single proprietary-style alarm dataset without a verified ground-truth causal graph; results are judged partly by expert knowledge.
- Precision on real data is lower than THP, PCMCI and ADM4, and SHD is higher than THP.
- Requires assuming an acyclic causal structure and sparsity via regularization weights that are hand-tuned.
Experimental setup
Datasets: PCIC 2021 simulated generation process Metropolitan cellular network alarm data (PCIC 2021)
Benchmarks: Simulated event sequences (varied alpha, mu, Delta, sample size, event types, nodes) Metropolitan cellular network alarm data
Metrics: Precision Recall F1 SHD SID
Baselines: PCMCI ACD MLE_SGL ADM4 THP
Builds on: topological Hawkes process amortized causal discovery NOTEARS acyclicity constraint Gumbel-Softmax
Closest papers by meaning
- A Simple yet Scalable Granger Causal Structural Learning Approach for Topological Event Sequences (NeurIPS 2024 · similarity 0.89)
- Exploring Neural Granger Causality with xLSTMs: Unveiling Temporal Dependencies in Complex Data (NeurIPS 2025 · similarity 0.88)
- Causal Discovery from Event Sequences by Local Cause-Effect Attribution (NeurIPS 2024 · similarity 0.88)
- Jacobian Regularizer-based Neural Granger Causality (ICML 2024 · similarity 0.87)
- CUTS+: High-Dimensional Causal Discovery from Irregular Time-Series (AAAI 2024 · similarity 0.87)
- Tangent Space Causal Inference: Leveraging Vector Fields for Causal Discovery in Dynamical Systems (NeurIPS 2024 · similarity 0.87)
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.