In order to extract more effective user tags from the attention behaviors of Internet users, attention characteristics of user behaviors were explored, and tags were predicted to improve the user portrait system. By constructing the adjacency matrix of the user and the attention behavior object, singular value decomposition was then performed to obtain the behavior feature matrix, and the logistic regression model was finally used to train the feature matrix and to predict the user's industry label. Experiments with 673 144 user behavior data on Weibo were carried out. The results show that the maximum prediction accuracy of 15 industry labels can reach 0.657 by using feature matrix of the user attention behavior to predict industry labels. The innovation is to alleviate the sparseness of user attention behaviors and improve the prediction effect of industry labels. The defects of users’ important labels can not be well reflected by improving user attention behaviors. The research provides a reference for the portrait labeling system of Internet users.