A New Progressive Multisource Domain Adaptation Network With Weighted Decision Fusion

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Multisource unsupervised domain adaptation (MUDA) is crucial for target classification using labeled source data.
• Aligning multiple source domains with a target domain in a common feature space presents significant challenges.

Purpose of the Study:

• To propose a novel progressive multisource domain adaptation network (PMSDAN) for improved target classification performance.
• To address the difficulties in mapping diverse source domains and a target domain into a unified feature space.

Main Methods:

• PMSDAN integrates multiple source domains into an auxiliary domain for initial distribution matching with the target domain.
• It then aligns the target domain's distribution separately with each source domain.
• A weighted hybrid maximum mean discrepancy (WHMMD) is proposed for optimization, considering interclass and intraclass discrepancies.

Main Results:

• The proposed PMSDAN demonstrates superior classification performance compared to existing state-of-the-art methods.
• Integrating source domains reduces discrepancies between source and target domains, and among source domains themselves.
• Separate alignment and weighted fusion effectively leverage knowledge from individual source domains.

Conclusions:

• The progressive approach of PMSDAN effectively improves target classification in multisource unsupervised domain adaptation scenarios.
• WHMMD offers a robust optimization strategy by considering both interclass and intraclass variations.
• PMSDAN provides a promising solution for leveraging heterogeneous labeled data in domain adaptation tasks.