Heterogeneous domain adaptation by class centroid matching and local discriminative structure preservation

Heterogeneous domain adaptation (HDA) aims at facilitating the target model training by leveraging knowledge from the heterogeneous source domain. HDA is a challenging problem since the domains are not consistent in not only data distribution but also feature space. Most HDA methods attend to search for a subspace, where the features and the distributions across domains can be aligned. However, these methods barely consider the shared semantic label space of two domains and do not align the decision boundaries of the two domains, which may cause misclassification. To address the above issue, we propose a novel HDA method called Class centroid Matching and local Discriminative structure Preservation (CMDP), which can transfer discriminative semantic source knowledge to the target domain. Specifically, we project cross-domain samples to regress the label matrix to align the discriminative directions of two domains. Then, we introduce the inner product strategy to align the distance and angle of the class centroids across domains, such that the discriminative source knowledge can more sufficiently transfer to the target domain. Besides, to further improve the quality of the class centroids in each domain, we propose a novel cross-domain graph embedding strategy to exploit the structure information of data more thoroughly. A simple and efficient optimization algorithm is designed to solve the CMDP model. Extensive experiments on heterogeneous datasets validate the superiority of our proposal over several advanced methods.