EFFICIENT AND MULTIPLY ROBUST RISK ESTIMATION UNDER GENERAL FORMS OF DATASET SHIFT

Area of Science:

• Statistical Machine Learning
• Data Science
• Causal Inference

Background:

• Machine learning models often suffer from limited target population data.
• Auxiliary data from related populations can mitigate this data scarcity.
• Existing domain adaptation and transfer learning methods have limitations in efficient risk evaluation.

Purpose of the Study:

• To develop efficient estimators for target population risk under various dataset shift conditions.
• To address the challenge of limited data in statistical machine learning.
• To improve the accuracy of risk evaluation in target domains using auxiliary data.

Main Methods:

• Leveraging semiparametric efficiency theory for risk estimation.
• Developing efficient and multiply robust estimators.
• Considering a general class of dataset shift conditions, including covariate, label, and concept shift.
• Allowing for partially nonoverlapping support between source and target populations.

Main Results:

• Efficient estimators for target population risk were developed.
• A straightforward specification test for dataset shift conditions was created.
• Efficiency bounds were derived for posterior drift and location-scale shift.
• Simulation studies confirmed efficiency gains from utilizing dataset shift conditions.

Conclusions:

• The proposed methods offer significant efficiency gains for risk estimation under dataset shift.
• The developed techniques enhance the utility of auxiliary data in machine learning.
• This work provides a robust framework for addressing data scarcity and domain adaptation challenges.