Domain Adaptive Bootstrap Aggregating

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Predictive algorithm performance degrades with distributional shifts between training and testing data (domain adaptation problem).
• Bootstrap aggregating (bagging) enhances algorithm stability, reduces variance, and prevents overfitting.

Purpose of the Study:

• To propose a novel domain adaptive bagging method to address the domain adaptation problem.
• To enhance the stability and accuracy of predictive algorithms in the presence of distributional shifts.

Main Methods:

• Developed a domain adaptive bagging method integrated with an iterative nearest neighbor sampler.
• Bootstrap samples are drawn to match the distribution of new testing data.
• Modified the method to accommodate anomalous samples (outliers) in test data.

Main Results:

• The proposed approach offers a general ensemble framework applicable to various classifiers and complex domains, including manifolds.
• Demonstrated effectiveness through theoretical support, simulations, and real-data applications.
• Successfully adapted algorithms to distributional shifts and handled anomalous test data.

Conclusions:

• The domain adaptive bagging method effectively mitigates performance degradation due to distributional shifts.
• The iterative nearest neighbor sampler is key to aligning bootstrap sample distributions with test data.
• This framework provides a robust solution for domain adaptation in machine learning.