CausalCompass Benchmark Tests Time-Series Causal Discovery Robustness
Researchers have introduced CausalCompass, a new benchmark framework aimed at evaluating the resilience of time-series causal discovery (TSCD) methods when their underlying assumptions are not met. This initiative addresses the limitations of existing benchmarks that rely on assumptions that can't be tested. In thorough assessments of several TSCD algorithms under eight different scenarios of assumption violation, no one method stood out as the best in every case. However, deep learning approaches generally performed well across multiple situations. The research also examines how sensitive these methods are to hyperparameters. You can find this study on arXiv with the identifier 2602.07915.
Key facts
- CausalCompass is a benchmark framework for time-series causal discovery robustness.
- It evaluates methods under violations of modeling assumptions.
- Eight assumption-violation scenarios were tested.
- No single method consistently performed best across all settings.
- Deep learning-based methods showed superior overall performance.
- The study includes hyperparameter sensitivity analysis.
- Published on arXiv with ID 2602.07915.
- Causal discovery from time series is a fundamental machine learning task.
Entities
Institutions
- arXiv