Abstract: How to derive the users’ hidden profiles using social relationships is studied. Considering the network structure of social network and characteristics of users’ data, the graph based semi-supervised learning algorithm is chose to predict users’ profiles. To improve the prediction accuracy, the attribute affinity is proposed to evaluate whether the value of an attribute is easy to be predicated, and different weight computing formulas are designed to calculate the relationship between users. The experimental data is collected from “renren network” and two attributes, hobbies and schools, are predicted in the experiments. The experimental results show that the strategies for computing weights among users are effective.