Integrating LDA into the Weighted Average Method for Semantic Friend Recommendation

Friend recommendation is a fundamental service for both social networks and practical applications. The majority of existing friend-recommendation methods utilize user profiles, social relationships, or static post content data, but rarely consider the semantic intentions and dynamic behaviors of users. In this paper, we propose FRec++, a friend recommendation method based on semantic and dynamic information. In FRec++, the first plus stands for new friend recommendation using a semantic model based on LDA while the next plus stands for the use of dynamic user attributes (e.g., behaviors or positions). More specifically, we first use the LDA method to generate semantic topics for user interests, and then compute the topic similarities between the target user and candidate friends. Next, we calculate the similarities of dynamic behaviors (i.e., forwarding, making comments, liking, and replying) and investigate the static attributes of users to measure the relevance of preferences. Finally, the weighted average method is used to integrate the above factors. We conducted experiments on the Weibo dataset and the results show that FRec++ outperforms several existing methods.