Research map / Probabilistic and scientific modeling / Causal discovery

CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series

ICLR 2025 · Probabilistic and scientific modeling › Causal discovery · approach: Causal discovery from 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.

Introduces CausalRivers, a large in-the-wild time-series causal discovery benchmark built from German river discharge data with two ground-truth graphs, and evaluates classical, deep, and naive baseline causal discovery methods on sampled subgraphs.

Problem

Causal discovery from time-series is evaluated mostly on synthetic data or small real-world examples; no large-scale, realistic time-series benchmark with ground-truth causal graphs exists.

Approach

Builds causal ground-truth graphs for river discharge measurement stations (over 1000 nodes) from agency metadata, Wikipedia, remote sensing and manual checks, then samples thousands of subgraphs with controlled structural characteristics (close, root cause, confounder, disjoint) and proposes simple naive baselines (CC, RP and restricted variants) that exploit river-specific priors.

Key results

• Naive baselines often match established causal discovery methods; RP+N achieves 0.81 AUROC on Root cause graphs with 3 nodes, outperforming all other methods on that set.
• VAR and Varlingam are the most robust methods across graph sets, with VAR reaching 0.81 AUROC on Close-3 and 0.82 on Random-3 subgraphs.
• Fine-tuning CP on RiversBavaria improves CP on all graph sets and outperforms the best Experiment Set 1 strategy on 50% of the datasets.

Measured effects reported in the paper

Surprising findings

• Simple naive baselines based on cross-correlation and river magnitude often perform similarly to sophisticated causal discovery methods.
• Deep learning methods CP and CDMI show no evidence of superiority over linear approaches on CausalRivers.

Limitations

• Ground truth graphs are constructed semi-automatically from Wikipedia, metadata and remote sensing, which may contain matching errors despite manual checks.
• Hyperparameter selection uses best-performing configuration per method with labels, which may overestimate performance.
• Limited preprocessing and no extensive hyperparameter search; results depend on default parameters.

Experimental setup

Datasets: RiversEastGermany RiversBavaria RiversElbeFlood

Benchmarks: CausalRivers subgraph sets (Close, Root cause, Random +1, Confounder, Random, Disjoint)

Metrics: AUROC F1 max max accuracy

Baselines: PCMCI VARLiNGAM DYNOTEARS VAR CDMI CP (Transformer) CP (GRU) RP CC Null model

Builds on: PCMCI VARLiNGAM DYNOTEARS Causal Pretraining CDMI

Closest papers by meaning

• CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · similarity 0.90)
• SPACETIME: Causal Discovery from Non-Stationary Time Series (AAAI 2025 · similarity 0.90)
• Arrow: Accelerator for Time Series Causal Discovery with Time Weaving (ICML 2025 · similarity 0.90)
• CausalTime: Realistically Generated Time-series for Benchmarking of Causal Discovery (ICLR 2024 · similarity 0.89)
• Discovering Mixtures of Structural Causal Models from Time Series Data (ICML 2024 · similarity 0.89)
• CausalDynamics: A large‐scale benchmark for structural discovery of dynamical causal models (NeurIPS 2025 · similarity 0.89)