Multi-domain Causal Structure Learning in Linear Systems.

Area of Science:

• Causal inference
• Machine learning
• Statistical modeling

Background:

• Causal structure learning aims to uncover cause-effect relationships from data.
• Existing methods often assume invariance across data domains, limiting applicability.
• Variations in causal parameters or noise distributions across domains pose challenges.

Purpose of the Study:

• To develop a general framework for causal structure learning in linear systems with multi-domain observational data.
• To address scenarios where causal coefficients or noise distributions vary across domains.
• To propose efficient methods for identifying causal direction in complex networks.

Main Methods:

• Leveraging the principle of independent changes in causal modules and parameters across domains.
• Developing methods for causal direction identification in two-variable systems.
• Generalizing these methods for causal structure learning in networks of variables.

Main Results:

• The proposed methods can identify causal direction with fewer than ten domains.
• When invariance holds, causal direction identification is typically achieved with just two domains.
• The approach unifies and generalizes previous multi-domain structure learning techniques.

Conclusions:

• The developed methods offer a robust approach to causal structure learning under domain variations.
• This work expands the applicability of causal inference techniques to more complex, real-world scenarios.
• Efficient identification of causal direction is achievable even without strict invariance assumptions.