Behind Distribution Shift: Mining Driving Forces of Changes and Causal Arrows.

Area of Science:

• Causal inference
• Machine learning
• Statistical modeling

Background:

• Causal discovery from nonstationary or heterogeneous data presents significant challenges.
• Existing methods often struggle with parameter changes over time or across datasets.

Purpose of the Study:

• To develop efficient methods for estimating the driving forces of nonstationarity in causal mechanisms.
• To introduce a novel approach for determining causal arrow directions using measures of dependence between changing causal modules.

Main Methods:

• Novel kernel embedding of nonstationary conditional distributions, avoiding sliding windows.
• Development of a dependence measure between changes in causal modules.

Main Results:

• Successfully extracted low-dimensional, interpretable representations of nonstationarity drivers.
• Demonstrated the effectiveness of the methods in determining causal arrow directions.
• Validated the approach on both synthetic and real-world datasets.

Conclusions:

• The proposed kernel embedding method offers an efficient way to analyze nonstationary causal systems.
• The developed dependence measure enhances causal discovery by leveraging changes in causal modules.
• This work advances causal discovery techniques for dynamic and heterogeneous data environments.