Aggregating Randomized Clustering-Promoting Invariant Projections for Domain Adaptation.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) uses labeled source data for unlabeled target data.
• Existing methods often focus on distribution divergence, neglecting intra-domain structure and data imbalance.
• This can hinder accurate target label inference in UDA.

Purpose of the Study:

• To propose a domain-invariant projection ensemble approach for UDA.
• To address both cross-domain distribution divergence and intra-domain structure issues.
• To improve the performance of target label inference in UDA.

Main Methods:

• A novel relaxed domain-irrelevant clustering-promoting term is used to find optimal projections.
• This term bridges the cross-domain semantic gap and enhances intra-class compactness.
• A 'sampling-and-fusion' framework with ensemble learning (e.g., majority voting) is employed for target data classification.

Main Results:

• The proposed method effectively bridges the cross-domain semantic gap.
• It increases intra-class compactness within both source and target domains.
• Experimental results show significant performance gains over state-of-the-art UDA methods on six visual benchmarks.

Conclusions:

• The domain-invariant projection ensemble approach is effective for UDA.
• It successfully tackles limitations of previous methods by considering intra-domain structure.
• The method demonstrates superior performance in object, face, and digit image recognition tasks.