Causal Discovery from Nonstationary/Heterogeneous Data: Skeleton Estimation and Orientation Determination.

Area of Science:

• Causal Inference
• Machine Learning
• Data Science

Background:

• Nonstationary and heterogeneous data are common, posing challenges for traditional causal discovery.
• Distribution shifts in data require advanced methods for accurate causal analysis.

Purpose of the Study:

• To develop a principled framework for causal discovery from nonstationary/heterogeneous data.
• To address challenges posed by changing data distributions in causal analysis.

Main Methods:

• Introduced Constraint-based causal Discovery from Nonstationary/heterogeneous Data (CD-NOD).
• Proposed an enhanced constraint-based procedure to detect local mechanism changes and recover causal structure skeletons.
• Developed a method for causal orientation using independence changes implied by distribution shifts.

Main Results:

• Successfully detected variables with changing local mechanisms.
• Recovered the skeleton of causal structures from observed variables.
• Determined causal orientations by leveraging distribution shift information.
• Demonstrated efficacy on synthetic and real-world datasets.

Conclusions:

• The CD-NOD framework provides an effective approach for causal discovery in the presence of nonstationary or heterogeneous data.
• Utilizing changes in data distributions is key to uncovering causal relationships in dynamic environments.