Domain Adaptation with Conditional Transferable Components.

Area of Science:

• Machine Learning
• Causal Inference
• Domain Adaptation

Background:

• Domain adaptation addresses differing data distributions between training and test sets.
• Existing methods often assume invariant conditional distributions P(Y|X), which may not hold.
• Previous work focuses on invariant components, overlooking identifiable changes in transferable components.

Purpose of the Study:

• To develop a domain adaptation method for causal systems where P(X|Y) and P(Y) change.
• To extract conditional transferable components with invariant conditional distributions after location-scale transformations.
• To simultaneously identify changes in P(Y) between domains.

Main Methods:

• Focuses on causal systems where Y is the cause of X.
• Extracts conditional transferable components invariant to location-scale (LS) transformations.
• Simultaneously identifies changes in the target distribution P(Y).

Main Results:

• Demonstrates theoretical effectiveness of the proposed method.
• Provides empirical validation on synthetic and real-world datasets.
• Shows improved domain adaptation performance in causal systems with shifting distributions.

Conclusions:

• The proposed method effectively handles domain adaptation in causal systems with changing P(X|Y) and P(Y).
• Extracting LS-invariant conditional transferable components is a viable strategy.
• The approach offers a robust way to adapt models when underlying causal mechanisms shift.