Domain adaptation based on the measure of kernel-product maximum mean discrepancy

Transfer learning is an important branch of machine learning, focusing on applying what has been learned in the old field to new problems. Maximum mean discrepancy (MMD) is used in most existing works to measure the difference between two distributions by applying a single kernel. Recent works exploit linear combination of multiple kernels and need to learn the weight of each kernel. Because of the singleness of single-kernel and the complexity of multiple-kernel, we propose a novel kernel-product maximum mean discrepancy (DA-KPMMD) approach. We choose the product of linear kernel and Gaussian kernel as the new kernel. Specifically, we reduce differences in the marginal and conditional distribution simultaneously between source and target domain by adaptively adjusting the importance of the two distributions. Further, the within-class distance is minimizing to differentiate samples of different classes. We conduct cross-domain classification experiments on three image datasets and experimental results show the superiority of DA-KPMMD compared with several domain adaptation methods. CCS CONCEPTS • Computing methodologies • Machine learning • Machine learning approaches • Kernel methods