Reducing bias in source-free unsupervised domain adaptation for regression.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Source-Free Unsupervised Domain Adaptation (SFUDA) aims to adapt models to new data without access to original training data, crucial for privacy.
• Existing SFUDA methods often struggle with theoretical guarantees for bias reduction, especially in regression tasks.

Purpose of the Study:

• To analyze the generalization error bounds of pseudo-label-based SFUDA methods in regression.
• To identify and address limitations in current SFUDA approaches for regression tasks.

Main Methods:

• Theoretical analysis of generalization error, identifying feature misalignment bias as a key factor.
• Proposal of a Bias-Reduced Regression (BRR) method incorporating Feature Distribution Alignment (FDA) and Hybrid Reliability Evaluation (HRE).

Main Results:

• Theoretical insights reveal that feature misalignment bias significantly impacts SFUDA performance in regression.
• The proposed BRR method, utilizing FDA and HRE, effectively reduces various biases.
• Experimental results show superior performance of BRR for SFUDA in regression tasks.

Conclusions:

• Feature misalignment bias is a critical, often overlooked, factor in SFUDA for regression.
• The BRR method offers a robust solution for SFUDA regression by systematically addressing multiple bias sources.
• The proposed techniques enhance the reliability and accuracy of domain adaptation in regression settings.