Research map / Probabilistic and scientific modeling / Causal discovery
Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed 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 REDCLIFF-S, a conditionally weighted superposition of cMLP-based Granger causal factor models with a supervised state model, to generate hypotheses of dynamic, nonlinear causal graphs in neural time series.
Problem
Hypothesis generation of causal graphs in neuroscience time series is limited by methods that assume static causal graphs or linear dynamics, failing to capture nonlinear, state-dependent brain interactions.
Approach
Model the dynamic causal graph as a state-conditioned weighted sum of static factor graphs, where each factor is a nonlinear cMLP forecaster and a state model produces factor scores (optionally supervised by behavioral labels via an MSE term), trained with sparsity, cosine-dissimilarity, and forecasting losses.
Key results
- On 12-11-2 synthetic systems, REDCLIFF-S achieved optimal F1 0.383±0.038 vs Regime-PCMCI 0.407±0.027 and LASAR 0.344±0.034, with the best mean comparative placement (2.0) across metrics.
- REDCLIFF-S improved over baselines' pairwise optimal F1 by roughly 22-28% on average in some experiments, with some improvements exceeding 60%, per the abstract.
- On D4IC HSNR, REDCLIFF-S (cMLP) achieved the highest average optimal F1 with a statistically significant margin, with ROC-AUC 0.629±0.018 vs cMLP-v2 0.567±0.015.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use REDCLIFF-S instead of Regime-PCMCI | 12-11-2 Synthetic Systems · 12-11-2 Synthetic Systems | ROC-AUC: 0.673±0.026 -> 0.705±0.022 | Helped |
| Use REDCLIFF-S instead of Regime-PCMCI | 12-11-2 Synthetic Systems · 12-11-2 Synthetic Systems | optimal F1: 0.407±0.027 -> 0.383±0.038 | Hurt |
| Use REDCLIFF-S (cMLP) instead of the strongest listed baseline | Synthetic System 6-2-2 · Synthetic System 6-2-2 | ROC-AUC: 0.612±0.050 -> 0.797±0.036 | Helped |
| Use REDCLIFF-S (cMLP) instead of the strongest listed baseline | Synthetic System 6-4-2 · Synthetic System 6-4-2 | ROC-AUC: 0.591±0.021 -> 0.682±0.037 | Helped |
| Use REDCLIFF-S (cMLP) instead of the strongest listed baseline | Synthetic System 12-11-2 · Synthetic System 12-11-2 | ROC-AUC: 0.572±0.020 -> 0.756±0.019 | Helped |
| Use REDCLIFF-S (cMLP) instead of the strongest listed baseline | Synthetic System 12-11-5 · Synthetic System 12-11-5 | ROC-AUC: 0.562±0.015 -> 0.743±0.019 | Helped |
| Use REDCLIFF-S (cMLP) instead of cMLP-v2 | D4IC LSNR · D4IC LSNR | ROC-AUC: 0.550±0.016 -> 0.632±0.014 | Helped |
| Use REDCLIFF-S (cMLP) instead of cMLP-v2 | D4IC MSNR · D4IC MSNR | ROC-AUC: 0.594±0.017 -> 0.635±0.016 | Helped |
| Use REDCLIFF-S (cMLP) instead of cMLP-v2 | D4IC HSNR · D4IC HSNR | ROC-AUC: 0.567±0.015 -> 0.629±0.018 | Helped |
| Remove the cosine similarity penalty (ρ = 0) (ablation) | REDCLIFF-S · Synthetic System 6-2-2; 6-4-2; 12-11-2; 12-11-5; D4IC HSNR | mean optimal F1 | Mixed |
| Use one factor (n_k = 1) (ablation) | REDCLIFF-S · Synthetic System 6-2-2; 6-4-2; 12-11-2; 12-11-5; D4IC HSNR | mean optimal F1 | Hurt |
| Set factor weights α = 1 (ablation) | REDCLIFF-S · Synthetic System 6-2-2; 6-4-2; 12-11-2; 12-11-5; D4IC HSNR | mean optimal F1 | Hurt |
Surprising findings
- Single-factor baselines attain fairly competitive F1 in some cases, seemingly learning an average factor graph.
- Regime-PCMCI and SLARAC trade off between F1/SHD and AID metrics, whereas REDCLIFF-S does not show this trade-off.
Limitations
- Hyperparameter tuning was costly and limited to a single repeat of one system, possibly limiting measured performance.
- No established nonlinear version of DeltaCon0 for evaluating functional similarity of nonlinear graphs; cosine penalty used as proxy.
- Nonlinear dynamic causal systems are theoretically not identifiable in general.
Experimental setup
Datasets: Synthetic VAR-based systems (6-2-2, 6-4-2, 12-11-2, 12-11-5) DREAM4 In Silico-Combined (D4IC) Tail Suspension Test (TST) LFP dataset Social Preference dataset
Benchmarks: Synthetic Systems 6-2-2, 6-4-2, 12-11-2, 12-11-5 D4IC LSNR/MSNR/HSNR TST case study
Metrics: optimal F1-score ROC-AUC Adjustment Identification Distance (parent AID error) Structural Hamming Distance comparative placement
Baselines: cMLP cLSTM DGCNN dCSFA-NMF DYNOTEARS NAVAR-R NAVAR-P Regime-PCMCI SLARAC QRBS LASAR
Builds on: cMLP DYNOTEARS NAVAR Regime-PCMCI dCSFA-NMF
Closest papers by meaning
- Learning dynamic representations of the functional connectome in neurobiological networks (ICLR 2024 · similarity 0.89)
- CausalDynamics: A large‐scale benchmark for structural discovery of dynamical causal models (NeurIPS 2025 · similarity 0.89)
- Identifying Neural Dynamics Using Interventional State Space Models (ICML 2025 · similarity 0.89)
- UnCLe: Towards Scalable Dynamic Causal Discovery in Non-linear Temporal Systems (NeurIPS 2025 · similarity 0.88)
- Discovering Mixtures of Structural Causal Models from Time Series Data (ICML 2024 · similarity 0.88)
- Learning Interpretable Hierarchical Dynamical Systems Models from Time Series Data (ICLR 2025 · 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.