Cross-Domain Pattern Classification With Distribution Adaptation Based on Evidence Theory.

Area of Science:

• Computer Science
• Machine Learning
• Pattern Recognition

Background:

• Domain adaptation (DA) is crucial when labeled data is scarce in the target domain.
• Existing DA methods transfer knowledge from labeled source domains to unlabeled target domains.
• Easily classified target objects can offer valuable information for classifying harder ones.

Purpose of the Study:

• To introduce a novel domain adaptation method, Distribution Adaptation based on Evidence Theory (DAET).
• To enhance cross-domain classification accuracy by leveraging complementary information from source and target domains.
• To effectively handle uncertainty in information sources during knowledge transfer.

Main Methods:

• DAET categorizes target objects into easy-target and hard-target groups.
• It generates classification results for hard-target objects using source domain data and easy-target object pseudo-labels.
• Weights for these results are estimated based on information source reliability (mean difference), and discounted using evidence theory for final decision-making.

Main Results:

• DAET combines information from both source and target domains.
• It utilizes evidence theory to manage uncertainty and discount classification results.
• Experimental results demonstrate significant improvements in classification accuracy compared to advanced DA methods.

Conclusions:

• DAET offers an effective approach for domain adaptation in pattern classification.
• The method successfully integrates complementary information and handles data uncertainty.
• DAET significantly enhances cross-domain classification performance.