Domain adaptation learning is an effective method for leveraging knowledge from the source domain with labels to a target domain without any labels. However, most prior methods have neglected the contribution of each sample to the integral measure when adaptively matching either marginal distribution, conditional distribution, or both between domains. In this article, an improved algorithm based on mean discrepancy embedding with structural similarity is proposed, which aims at the contribution of each sample to the integral measure on the performance of target domain learning model by using labeled source samples and unlabeled target samples. The discrepancy between both marginal and conditional distribution are minimized with dimensionality reduction procedure to feature extraction with structural similarity weights for all samples from the source and target domains. The results of empirical analysis demonstrate that the proposed method has better performance over several state-of-the-art methods in credit risk classification.`