An Evidential Multi-Target Domain Adaptation Method Based on Weighted Fusion for Cross-Domain Pattern Classification

Area of Science:

• Machine Learning
• Computer Vision
• Pattern Recognition

Background:

• Cross-domain pattern classification often relies on labeled source data to classify unlabeled target data.
• Multiple target domains present an opportunity to improve classification by utilizing high-confidence predictions as pseudo-supervised information.

Purpose of the Study:

• To propose an evidential multi-target domain adaptation (EMDA) method for effectively utilizing information from single-source and multiple target domains.
• To enhance classification performance in target domains by integrating pseudo-supervised information from different target domains.

Main Methods:

• EMDA aligns source and target domain distributions using Maximum Mean Discrepancy (MMD) and covariance difference reduction.
• A novel approach combines initial predictions with pseudo-supervised classifier outputs using evidence theory.
• Weighting factors are estimated based on cross-domain distribution discrepancies to discount and fuse soft classification results.

Main Results:

• The proposed EMDA method demonstrates improved performance compared to existing domain adaptation techniques.
• Effective utilization of pseudo-supervised information from multiple target domains leads to better classification accuracy.
• The evidential combination strategy, incorporating estimated weighting factors, enhances the reliability of the final classification decision.

Conclusions:

• EMDA offers a robust framework for cross-domain pattern classification by effectively integrating information from multiple target domains.
• The method's ability to leverage pseudo-supervised signals and employ evidence theory for fusion represents a significant advancement.
• Empirical validation on benchmark datasets confirms the superiority of EMDA over advanced domain adaptation methods.