Among a lot of the multi-view dimension reduction methods, there are few methods for semi-supervised problem since it is difficult in obtaining the sample labels. In this paper, a new semi-supervised dimensionality reduction method which is called multi-view semi-supervised similarity projection (MSSP) is proposed. A new metric of similarity and correlations between views is firstly defined and then a regularization term based on the new metric is proposed. MSSP finds the low-dimensional projection matrix by maximizing a discriminant term and the new regularization term for semi-supervised data. Compared with other methods, MSSP makes full use of all sample information; MSSP emphasizes the samples that have different local structure from different views should be more similar. 1NN is finally used to test the effectiveness of MSSP. Experiments on YALE and ORL data sets show that MSSP is superior to other methods.