Heterogeneous Domain Adaptation With Generalized Similarity and Dissimilarity Regularization

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Heterogeneous domain adaptation (HDA) addresses transfer learning with differing feature representations between source and target domains.
• Existing matrix factorization HDA methods effectively learn transferable features but neglect label information for cross-domain sample similarity and separability.
• This limitation leads to cross-domain bias and potential class mixing in the common subspace, hindering discriminative target feature learning.

Purpose of the Study:

• To propose a novel matrix factorization-based HDA method, HGSDR (HDA with generalized similarity and dissimilarity regularization).
• To improve cross-domain feature learning by incorporating label information to explore sample similarity and separability.
• To enhance the discriminative power of target domain features by mitigating cross-domain bias.

Main Methods:

• Developed HGSDR, a matrix factorization-based HDA method incorporating generalized similarity and dissimilarity regularization.
• Introduced a similarity regularizer using a cross-domain Laplacian graph with label information to capture similarities between identical cross-domain classes.
• Proposed a dissimilarity regularizer based on inner products to increase separability between different cross-domain classes, while preserving neighbor relationships for unlabeled target samples.

Main Results:

• HGSDR effectively matches domain distributions globally and at the sample level.
• The method learns more discriminative features for target samples compared to existing approaches.
• Extensive experiments on benchmark datasets validated the superiority of HGSDR over state-of-the-art methods.

Conclusions:

• HGSDR significantly improves heterogeneous domain adaptation by utilizing label information for enhanced similarity and dissimilarity regularization.
• The proposed method effectively reduces cross-domain bias and learns discriminative features for target domains.
• HGSDR offers a superior approach for tackling HDA challenges, particularly when dealing with heterogeneous features and limited labeled target data.