Correlated Matching and Structure Learning for Unsupervised Domain Adaptation

Abstract

AbstractFor cross-domain tasks in real-world, a source domain and a target domain often have different marginal probability distribution and conditional probability distribution. To leverage the distribution difference between the source and target domain, domain adaptation has been applied in many fields. Unfortunately, as most of the existing domain adaptation methods only focus on eliminating the distribution discrepancy between the two domains, they do not make full use of the correlation information and data distribution structure between the two domains. In this paper, we put forward a novel domain adaptation method named correlated matching and structure learning (CMSL), which considers the association information between source and target domains, and extracts the feature representation and thus can learn the maximization correlation features between the two domains. Simultaneously, the class centroids of the source data are used to cluster the target data, and a local manifold self-learning strategy is introduced to the target domain to preserve the underlying structure of the data. Experimental results on six data benchmarks show that our proposed method achieves good classification performance and outperforms several state-of-the-art unsupervised domain adaptation methods.KeywordsDomain adaptationCross-domain learningCorrelated matchingStructure learning