Scalable Causal Discovery in
Nonlinear Temporal Systems
with Meta-Learning
Learn across targets. Discover directed dependencies.
Scale to large temporal systems.
University of Tennessee at Chattanooga
Code repository link to be added.
Each target variable defines a prediction task.
Sample targets, adapt on support, update on query.
Read out continuous scores after training.
A quadratic problem.
A shared learning strategy.
SCDM treats the prediction of each target variable as a related task, sharing a temporal predictor and a causal-strength matrix across the system.
When the number of variables grows, so does the space of possible directed relations. Target-wise fitting and repeated conditional testing can become expensive. SCDM processes a subset of target tasks per episode, allowing information from those tasks to update a shared initialization.
Task sampling reduces per-episode target processing. The dense structural matrix and complete graph readout still depend on system dimension.
Counts exclude self-edges. This calculation illustrates graph size, not measured runtime.
Read the abstract
Adapt locally.
Learn collectively.
Temporary task adaptation and persistent shared learning play different roles. Follow one SCDM episode below.
Recover directed structure.
Across controlled systems.
Known directed graphs let us evaluate edge-ranking performance. Results vary with the dataset and the metric.
All benchmark results
Bold: highest reported meanValues transcribed from the manuscript’s controlled-recovery table. AUROC and AUPRC are threshold-independent and unitless; self-edges are excluded. Rankings use reported means and do not imply statistical significance.
From brain networks
to traffic systems.
Controlled simulations and real-world observations answer different questions. Explore the figures and their evaluation roles.
These are benchmark dimensions. The voxel simulator includes latent regional drivers; LargeST has no complete directed causal ground truth.
A recovery principle,
with explicit conditions.
For a fully observed delayed system, the graph can be recovered when estimation errors remain small relative to the population risk margin.
- εstat
- Finite-sample error under temporal dependence
- εmeta
- Residual from finite episodic meta-training
- εreadout
- Difference between implemented and ideal readout
Under the stated assumptions, a threshold inside the separating margin recovers the delayed graph with high probability.
The guarantee assumes causal sufficiency, a correct history order, temporal regularity, a positive population margin, and controlled estimation errors. It does not cover arbitrary hidden confounding. Predictive readout and geographic locality are diagnostics; neither alone verifies an individual causal effect.
The paper and
what comes next.
View publication ↗Reference this work.
Citation uses the title and author list from the OpenReview record.