Combination of Transferable Classification With Multisource Domain Adaptation Based on Evidential Reasoning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multisource domain adaptation leverages knowledge from multiple sources to improve target domain classification.
• Traditional methods may not effectively handle varying reliabilities of source domains.
• Evidential reasoning offers a robust framework for uncertainty management in classification.

Purpose of the Study:

• To propose a decision-level combination method for multisource domain adaptation.
• To enhance classification accuracy in the target domain by integrating complementary knowledge.
• To reduce classification errors through a cautious decision-making strategy.

Main Methods:

• A decision-level combination strategy based on evidential reasoning is proposed.
• Classification results from source domains are weighted by domain consistency and combined using Dempster's rule.
• A neighborhood-based cautious decision-making rule is developed to handle classification uncertainty.

Main Results:

• The proposed method effectively combines information from multiple source domains.
• The cautious decision-making rule successfully characterizes partial imprecision and reduces error risk.
• Experimental results demonstrate significant improvements in classification accuracy compared to existing methods.

Conclusions:

• The developed evidential reasoning-based multisource domain adaptation method improves classification performance.
• The cautious decision-making rule offers a beneficial trade-off between precision and error reduction.
• This approach provides a robust solution for pattern classification with multiple, potentially conflicting, data sources.