Domain adaptation from multiple sources: a domain-dependent regularization approach.

Area of Science:

• Machine Learning
• Computer Vision
• Information Retrieval

Background:

• Multiple source domain adaptation is crucial for leveraging labeled data across different domains.
• Existing methods often struggle with heterogeneous source domains and target domain generalization.

Purpose of the Study:

• To propose a novel framework, Domain Adaptation Machine (DAM), for multiple source domain adaptation.
• To develop two new methods, FastDAM and UniverDAM, enhancing target classifier robustness and efficiency.
• To improve label prediction accuracy in target domains using unlabeled data.

Main Methods:

• Developed a Domain Adaptation Machine (DAM) framework utilizing base classifiers.
• Introduced a domain-dependent regularizer based on smoothness assumption for target classifier learning.
• Proposed FastDAM incorporating LS-SVM and sparsity regularization for efficient prediction.
• Proposed UniverDAM utilizing source domain instances as Universum for enhanced generalization.

Main Results:

• FastDAM achieves fast label prediction by learning a sparse target classifier.
• UniverDAM enhances generalization by incorporating source domain instances as Universum.
• Both methods demonstrated superior performance over existing multiple source domain adaptation techniques.
• Evaluated on TRECIVD 2005 (video concept detection) and 20 newsgroups/email spam (document retrieval) datasets.

Conclusions:

• The proposed DAM framework and its variants, FastDAM and UniverDAM, offer effective solutions for multiple source domain adaptation.
• These methods significantly improve performance in challenging tasks like large-scale video concept detection and document retrieval.
• The domain-dependent regularizer and Universum approach contribute to more robust and generalizable target classifiers.