Domain Adaptation with Data Uncertainty Measure Based on Evidence Theory.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Domain adaptation seeks to leverage labeled source data for target tasks.
• Mismatched source and target domains introduce uncertainty, impacting classifier reliability.
• Existing methods often overlook this uncertainty, focusing solely on distribution matching.

Purpose of the Study:

• To address the limitations of current domain adaptation techniques.
• To incorporate source domain data uncertainty into adaptive classifier learning.
• To develop a more robust and accurate domain adaptation framework.

Main Methods:

• Utilized evidence theory to design an 'evidence net' for estimating source domain data uncertainty.
• Developed a general loss function incorporating the uncertainty measure for adaptive classifiers.
• Extended the proposed loss function to support vector machines (SVM).

Main Results:

• Demonstrated the effectiveness of the uncertainty measure in improving adaptive classifier performance.
• Achieved reliable and optimal classification results on target domain tasks.
• Validated the approach through numerical experiments on simulation datasets and real-world applications.

Conclusions:

• The proposed method effectively quantifies and utilizes source domain data uncertainty.
• The uncertainty-aware adaptive classifier offers improved performance over traditional methods.
• This work provides a novel perspective for enhancing domain adaptation strategies.