Research map / Probabilistic and scientific modeling / Causal discovery
Causal Discovery from Event Sequences by Local Cause-Effect Attribution
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 a cause-effect matching model for event sequences with dynamic delays and the CASCADE algorithm, which uses an MDL score grounded in the Algorithmic Markov Condition to recover causal DAGs over event types.
Problem
Causal discovery from multivariate event sequences, where Granger-causality-based methods cannot handle instantaneous effects and often return spurious edges.
Approach
Model each cause event as triggering at most one effect event with a random delay (or not at all), score graphs with an MDL instantiation of the Algorithmic Markov Condition, and build the DAG by adding edges in topological order, which yields identifiability for both instant and delayed effects.
Key results
- On the Network Alarms dataset CASCADE attains SHD 42 versus 127 for THP, 214 for NPHC, and 1564 for CAUSE, correctly identifying 142 of 147 causal edges.
- On instant-effect synthetic data CASCADE achieves unnormalized SHD 32.8 versus 46.85 for NPHC (second best).
- CASCADE recovers the true DAG without spurious edges on a no-structure sanity check, while THP reports at least one spurious edge in 45% of cases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CASCADE instead of the strongest reported baseline, NPHC | synthetic data with 90% instant effects · Synthetic data, 90% instant effects | SHD: 19.45 vs 47.50 | Helped |
| use CASCADE instead of the strongest reported baseline, NPHC | synthetic data with exclusively instant effects · Synthetic data, exclusively instant effects | SHD: 32.80 vs 46.85 | Helped |
| use CASCADE instead of THP | Network Alarms causal discovery · Network Alarms | SHD: 42 vs 127 | Helped |
| use CASCADE instead of NPHC | Network Alarms causal discovery · Network Alarms | SHD: 42 vs 214 | Helped |
| use CASCADE instead of CAUSE | Network Alarms causal discovery · Network Alarms | SHD: 42 vs 1564 | Helped |
| constrain the search to the known network edges | CASCADE · Network Alarms | causal edges correctly identified: 142 out of 147 | Helped |
| increase the number of colliders from 5 to 20 | CASCADE · Multiple colliders experiment | F1: 0.97 ±0.01 -> 0.82 ±0.01 | Hurt |
| vary the assumed delay distribution from exponential to other distributions (ablation) | CASCADE · Synthetic data, distribution misspecification | F1: 0.82; 0.81; 0.76; 0.75 | Mixed |
| increase the event-type count | CASCADE · Synthetic data, event-type scalability | F1 | Hurt |
| increase noise | CASCADE · Synthetic data, noise experiment | causal graph recovery | Hurt |
| increase expected events generated per cause in Hawkes data | CASCADE · Hawkes-process synthetic data | F1 | Mixed |
Surprising findings
- THP reports at least one spurious edge in 45% of sanity-check cases on data with no causal structure.
- MDLH did not terminate within 96 hours even for 15 variables.
Limitations
- Only one-to-one direct cause-effect matching is modeled; a single event causing multiple events or multiple events jointly causing one are excluded.
- Only excitatory effects are modeled, so absence of events due to a cause is not captured.
- Identifiability under instantaneous effects depends on the trigger and noise probability strengths, which are not operationalized in the search.
Experimental setup
Datasets: Network Alarms (NeurIPS 2023 CSL competition, Huawei) Global Banks daily return volatility dataset Daily Activities dataset synthetic event sequences from the authors' generator
Benchmarks: Network Alarms Global Banks Daily Activities synthetic event sequences
Metrics: SHD SID F1
Baselines: THP CAUSE NPHC MDLH
Builds on: Algorithmic Markov Condition minimum description length Hawkes processes Granger causality
Closest papers by meaning
- MOCHA: Discovering Multi-Order Dynamic Causal Structure in Temporal Point Processes (NeurIPS 2026 · similarity 0.89)
- CausalDynamics: A large‐scale benchmark for structural discovery of dynamical causal models (NeurIPS 2025 · similarity 0.88)
- Flow based approach for Dynamic Temporal Causal models with non-Gaussian or Heteroscedastic Noises (NeurIPS 2025 · similarity 0.88)
- CausalTime: Realistically Generated Time-series for Benchmarking of Causal Discovery (ICLR 2024 · similarity 0.88)
- TNPAR: Topological Neural Poisson Auto-Regressive Model for Learning Granger Causal Structure from Event Sequences (AAAI 2024 · similarity 0.88)
- Integrating Background Knowledge for Scalable Causal Discovery (NeurIPS 2026 · 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.