POTENTIAL DOMAIN UNSUPERVISED DOMAIN ADAPTIVE BASED ON MULTIPLE KERNEL LEARNING ALGORITHM

In order to improve the adaptive performance of unsupervised domain, a potential domain unsupervised domain adaptive method based on multiple kernel learning algorithm is proposed. Three criteria for potential domain discovery were proposed: the maximization of data compactness and significance in a single potential target domain, and the minimization of the total divergence from potential target domain to source domain. The projected source data in the potential feature space was regarded as different views of the source domain, which reduced the difference between the source domain and the specific potential target domain. The experimental results on different tasks show that the proposed algorithm has better classification accuracy and robustness.