In this paper, we adopt the Gabor transform to extract features of facial expression images and use a series of locally linear embedding(LLE) algorithms to reduce the data dimension. The LLE algorithm widely used in facial expression recognition is a kind of nonlinear dimension reduction algorithm. It is able to make dimension-reduced data keep the original topology. Because the LLE algorithm does not take the category information of samples into account, the supervised locally linear embedding(SLLE) algorithm appears. But the SLLE algorithm only considers the category information of samples, and does not take the relationship among various expressions into account. Therefore, in this paper, we propose an improved SLLE algorithm, which regards the neutral expression as the center of the other expressions. The results of facial expression recognition experiments on the JAFFE database show that our algorithm obtained better facial expression recognition rate compared with the LLE algorithm and the SLLE algorithm. Our algorithm is more effective