Researchers survey CI tests for causal AI discovery
33-page survey reveals how CI tests power next-gen causal AI systems
A team of researchers—Pavel Averin, Theodoros Moysiadis, and Ioannis Katakis—published a 33-page survey in *Transactions on Machine Learning Research* (July 2026) that dissects conditional independence (CI) tests, the statistical engine behind constraint-based causal discovery.
The survey categorizes widely used CI methods into six families: partial-correlation, contingency-table, regression, nearest-neighbor, kernel, and machine-learning-based. It examines how each method’s assumptions, robustness, and scalability influence causal graph reconstruction, particularly in high-dimensional biomedical data. The paper also links test-level properties (e.g., power decay with conditioning set size) to graph-level errors in skeleton recovery and v-structure orientation, while comparing adoption across major R and Python libraries.
- Survey benchmarks six CI testing families used in causal discovery algorithms like PC and FCI
- Focuses on robustness and scalability in high-dimensional, mixed-type biomedical datasets
- Highlights open challenges: mixed-type CI testing without discretization and small-sample error control
Why It Matters
Critical for advancing explainable AI and causal inference in healthcare and genomics