Research map / Probabilistic and scientific modeling / Causal discovery
UnCLe: Towards Scalable Dynamic Causal Discovery in Non-linear Temporal Systems
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 UnCLe, a parameter-sharing TCN Uncoupler/Recoupler autoencoder with auto-regressive dependency matrices that yields dynamic causal graphs via datapoint-wise perturbation error gains (and static graphs via aggregation), evaluated on synthetic static/dynamic benchmarks and MoCap/traffic data.
Problem
Discovering time-resolved (dynamic) causal graphs from observational nonlinear time series, where causal relations evolve over time, while scaling to many variables; prior neural Granger methods produce static graphs or scale poorly.
Approach
A parameter-sharing TCN autoencoder (Uncoupler/Recoupler) disentangles series into C semantic channels, a set of linear auto-regressive Dependency Matrices predicts future latents, and dynamic causal strength is the per-timestep increase in prediction error when a series is permuted in time.
Key results
- On TVSEM, UnCLe(P) reaches AUROC 1.000 and AUPRC 1.000 versus 0.733/0.400 for GVAR and 0.467 for the best static graph.
- On ND8, UnCLe(P) achieves AUROC 0.921 vs 0.723 for GVAR and 0.905 for the static best in AUROC, and surpasses the static best in AUROC.
- On Lorenz#1, UnCLe(P) achieves AUROC 0.999 and AUPRC 0.996, the top scores among the compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use temporal permutation instead of noise injection or zero-masking (ablation) | UnCLe(P) · Lorenz#1 | AUROC: .981±.056 / .974±.082 -> .999±.002 | Helped |
| Remove perturbation (ablation) | UnCLe(P) · Lorenz#1 | AUROC: .999±.002 -> .500±.000 | Hurt |
| Increase autoregressive lag from 1 to 2 or 4 (ablation) | UnCLe(P) · Lorenz#1 | AUROC: .999±.002 -> .961±.042 (lag 2); .877±.076 (lag 4) | Hurt |
| Remove parameter sharing (ablation) | UnCLe(P) · Lorenz#3 | causal discovery performance | Hurt |
| Remove Dependency Matrices and disable parameter sharing (ablation) | UnCLe(P) · Lorenz#3 | AUROC: below 0.5 | Hurt |
| Remove the prediction task, training only on reconstruction (ablation) | UnCLe(P) · Lorenz#3 | causal discovery performance | Hurt |
| Use static causal graphs rather than time-varying causal discovery | Static Best · TVSEM | AUROC: 0.467±.000 vs 1.000±.000 | Hurt |
| Use dynamic perturbation-based causal discovery instead of GVAR | UnCLe(P) · ND8 | AUROC: 0.723±.016 -> 0.921±.007 | Helped |
Surprising findings
- Disabling the Dependency Matrices drops AUROC below 0.5 (random), and no-perturbation gives AUROC 0.5.
- Increasing the autoregressive lag from 1 to 2 or 4 reduced performance for both UnCLe(P) and UnCLe(A).
Limitations
- No formal identifiability guarantees for the learned dynamic causal graph.
- Dynamic ground truth is available only for synthetic datasets; MoCap and traffic results lack quantitative dynamic validation beyond skeleton missing rate and visual inspection.
- Grid search on hyperparameters to maximize AUROC may favor the proposed method; assumes no hidden confounders.
Experimental setup
Datasets: Lorenz96 fMRI NC8 FINANCE TVSEM ND8 MoCap METR-LA PEMS-BAY
Benchmarks: Lorenz96 NC8 FINANCE TVSEM ND8 fMRI MoCap METR-LA PEMS-BAY
Metrics: AUROC AUPRC ACC missing rate
Baselines: VAR PCMCI cMLP TCDF GVAR VARLiNGAM DYNOTEARS CUTS+ JRNGC
Builds on: Granger causality neural Granger causality TCDF JRNGC CUTS+
Closest papers by meaning
- Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains (ICML 2025 · similarity 0.88)
- MOCHA: Discovering Multi-Order Dynamic Causal Structure in Temporal Point Processes (NeurIPS 2026 · similarity 0.88)
- CausalDynamics: A large‐scale benchmark for structural discovery of dynamical causal models (NeurIPS 2025 · similarity 0.88)
- Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series (ICML 2025 · similarity 0.88)
- Causal Discovery from Conditionally Stationary Time Series (ICML 2025 · similarity 0.88)
- CausalTime: Realistically Generated Time-series for Benchmarking of Causal Discovery (ICLR 2024 · similarity 0.88)
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.