Research map / Probabilistic and scientific modeling / Causal discovery
Towards the Disappearing Truth: Fine-Grained Joint Causal Influences Learning with Hidden Variable-Driven Causal Hypergraphs in Time Series
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 CHGNN, a hidden variable-driven causal hypergraph neural network that models fine-grained joint causal influences to recover weak Granger causalities in multivariate time series.
Problem
Causal discovery under the Granger framework with neural networks loses weak causalities (the "causality disappearing problem") because sparsity penalties shrink their weights to zero, and existing methods model joint causal influences only implicitly and coarsely.
Approach
Introduce hidden variables that define a directed causal hypergraph whose hyperedges represent fine-grained joint causal influences, combine a multi-level causal path with an information aggregation path (dual causal information transfer mechanism), and train with a multi-view collaborative optimization constraint (JS-divergence diversity loss, collection sparsity loss, forecasting loss).
Key results
- CHGNN reaches 93.10 AUROC / 90.14 AUPRC on Joint-VAR-10 and 96.03 AUROC / 88.91 AUPRC on VAR-30, outperforming all seven baselines in the reported table.
- On Lorenz-96 CHGNN achieves 92.54 AUROC and 82.41 AUPRC versus 91.34 / 52.45 for TCDF (best baseline AUPRC is GVAR at 56.96).
- Ablations show removing the information aggregation path, the JS diversity loss, the sparse loss, or hypergraph convolution each degrades causal discovery results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CHGNN instead of the strongest baseline | Causal discovery models · VAR-30 | AUPRC: 83.33 -> 88.91 | Helped |
| use CHGNN instead of the strongest baseline | Causal discovery models · Joint-VAR-10 | AUPRC: 83.66 -> 90.14 | Helped |
| use CHGNN instead of the strongest baseline | Causal discovery models · Joint-VAR-20 | AUPRC: 69.99 -> 80.81 | Helped |
| use CHGNN instead of the strongest baseline | Causal discovery models · Lorenz-96 | AUPRC: 99.83 -> 98.72 | Hurt |
| use CHGNN instead of the strongest baseline | Causal discovery models · fMRI | AUPRC: 56.96 -> 82.41 | Helped |
| add the information aggregation path (ablation) | CHGNN · VAR-30 and Joint-VAR-20 | causal discovery performance | Helped |
| add the JS divergence-guided diversity loss (ablation) | CHGNN · VAR-30 and Joint-VAR-20 | causal discovery performance | Helped |
| add the collection sparse loss (ablation) | CHGNN · VAR-30 and Joint-VAR-20 | causal discovery performance | Helped |
| use hypergraph convolutions instead of common graph convolutions (ablation) | CHGNN · VAR-30 and Joint-VAR-20 | causal discovery performance | Helped |
Limitations
- Evaluated only on synthetic benchmarks plus one real dataset; real-world gains are described qualitatively.
- Joint causal influences are modeled with hyperparameters (mu, lambda, k, eta) whose sensitivity is not discussed in the main text.
- Causal graph requires thresholding for binary evaluation, though AUROC/AUPRC avoid this.
Experimental setup
Datasets: VAR-30 Lorenz-96 fMRI Joint-VAR-10 Joint-VAR-20 gas turbine fault data (GTF)
Benchmarks: VAR-30 Lorenz-96 fMRI Joint-VAR-10 Joint-VAR-20
Metrics: AUROC AUPRC
Baselines: Lasso Granger VARLiNGAM TCDF eSRU GVAR cMLP cLSTM
Builds on: neural Granger causality (cMLP/cLSTM) hypergraph convolution graph convolutional networks
Closest papers by meaning
- CUTS+: High-Dimensional Causal Discovery from Irregular Time-Series (AAAI 2024 · similarity 0.89)
- UnCLe: Towards Scalable Dynamic Causal Discovery in Non-linear Temporal Systems (NeurIPS 2025 · similarity 0.87)
- Learning Causal Relations from Subsampled Time Series with Two Time-Slices (ICML 2024 · similarity 0.87)
- Discovering Mixtures of Structural Causal Models from Time Series Data (ICML 2024 · similarity 0.87)
- GCAD: Anomaly Detection in Multivariate Time Series from the Perspective of Granger Causality (AAAI 2025 · similarity 0.87)
- TNPAR: Topological Neural Poisson Auto-Regressive Model for Learning Granger Causal Structure from Event Sequences (AAAI 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.