Local Causal Discovery in Multiple Manipulated Datasets.

Area of Science:

• Causal inference
• Machine learning
• Statistical modeling

Background:

• Distinguishing direct causes from effects is crucial for understanding complex systems.
• Manipulated datasets offer richer causal information than observational data for causal structure learning.
• Real-world scenarios often involve multiple datasets with unknown manipulated variables and non-identical distributions.

Purpose of the Study:

• To propose a new algorithm for identifying the direct causes and direct effects of a target variable.
• To leverage the interventional properties of causal models across multiple manipulated datasets.
• To address the challenge of unknown manipulated variables and non-identical data distributions.

Main Methods:

• Utilizing a backward framework to learn the parents and children (PC) of a target variable from multiple manipulated datasets.
• Orienting edges connected to the target variable based on the assumption of no target manipulation.
• Orienting remaining undirected edges by identifying invariant V-structures across datasets.

Main Results:

• The proposed algorithm successfully identifies the local causal structure of a target variable.
• The method is effective even when manipulated variables are unknown and data distributions differ.
• Experimental validation on standard Bayesian networks confirms the algorithm's effectiveness.

Conclusions:

• The developed algorithm is the first to identify local causal structure from multiple manipulated datasets with unknown manipulated variables.
• The approach enhances causal discovery by effectively utilizing interventional data.
• This work contributes to more robust causal inference in complex, real-world settings.