REX: Causal discovery based on machine learning and explainability techniques

Explainability can improve causal discovery: REX demonstrates that incorporating explainability techniques (specifically Shapley values) into causal discovery methods can help identify which variables actually cause changes in other variables, rather than just correlating with them—an important distinction for understanding complex systems.

Strong performance on real-world data: When tested on actual biological data (the Sachs single-cell protein-signaling dataset), REX achieved 95.2% precision in identifying true causal relationships while avoiding false connections, suggesting practical applicability beyond theoretical exercises.

Works across different data types: REX performed well on diverse synthetic datasets including non-linear relationships and different noise patterns, indicating the method is robust and not limited to oversimplified scenarios—a key requirement for real-world deployment in healthcare, economics, and other domains.

Addresses a gap in current methods: The paper identifies that existing causal discovery approaches don't leverage explainability techniques, and REX fills this gap by combining machine learning with interpretability methods, potentially making causal discovery more transparent and trustworthy for decision-makers.

Minimizes false positives: By reducing incorrect causal relationships while accurately recovering true ones, REX could help prevent costly mistakes in high-stakes domains like healthcare where incorrect causal assumptions could lead to ineffective or harmful interventions.