Abstract: Sentiment analysis technique can help make effective decisions and solutions by automatically discriminating the sentiment polarity in a practical application scene. However, it requires a large amount of annotated samples. To reduce the dependence on manual annotation, some researchers propose the domain adaptation based cross-domain sentiment analysis methods, which transfer a source domain model trained on an adequately labeled dataset to an unlabeled target domain. However, existing domain adaptation methods transfer from only one angle, namely, reducing the discrepancy of domain-specific features or simply extracting the domain-invariant features. To make use of domain-specific features and domain-invariant features together, we propose an unsupervised domain adaptation sentiment analysis algorithm in this paper for unsupervised cross-domain sentiment classification tasks. The algorithm reduces the domain discrepancy on different semantic layers with a progressive transfer strategy, and adopts the synergistic optimization of domain adaptation algorithm in high-level semantic subspace to transfer the domain knowledge of cross-domain text data. We validate our algorithm on 2 public review datasets with 24 cross-domain sentiment classification tasks. It is compared with 4 types of domain adaptation algorithms. The results show that our algorithm achieves the highest average accuracy. Moreover, it has better performance than the existing unsupervised cross-domain text sentiment classification algorithms in terms of the performance of classification and transferring.