“Like Attracts Like!”– A Social Recommendation Framework Through Label Propagation

Recently label propagation recommendation receives much attention from both industrial and academic fields due to its low requirement of labeled training data and effective prediction. Previous methods propagate preferences on a user or item similarity graph for making recommendation. However, they still suffer some major problems, including data sparsity, lack of trustworthiness, cold-start problem. By observation, the currently booming social network has some characteristics to remedy these problems. (1) Most of the user connections in either social network or real life can inflect information about users’ interest similarity by “Like Attracts Like”, which can improve propagation graph construction. (2) Social connections can inflect trustworthiness information for user similarity, where connections are not built randomly but based on their trust. (3)Social network can provide user connection data as the supplementation of sparse ratings, which can also solve the cold-start problem when one new user has no rating history but social network. In order to improve the recommendation accuracy, we propose a social label propagation recommendation framework. In addition, we also construct the traditional user similarity graph for combination with social network to solve the noise and multi-interest problem in social network. Finally, we implement Green’s function semi-supervised learning algorithm for label propagation recommendation on the real world recommendation data. The empirical results demonstrate the effectiveness of our proposed social recommendation framework.