Microblog Friends Recommedation Using Heterogeneous Data

As a new social media, microblogging service plays an increasing role in people’s daily life. However, with the rapid growth of microblogging, a commercial microblog ging system usually has a large number of users and diverse information. So it becomes vitally important to push information of interest to each user from massive information in the microblogging system.In the field of personalized recommendation, research on e-commerce recommendation is mature, yet microblog friends recommendation is a relatively new topic. Existing microblog friends recommendation techniques are almost based on a single data source, the performance is very limited. In this paper we develop an algorithm framework to recommend microblog friends by exploiting the heterogeneous data in the microblogging system.In this paper we first studied recommendation in traditional E-commerce, we elaborated state-of-the-art collaborative filtering (CF) techniques, including neighbour based CF and model based CF. Then we studied two microblog friends recommendation techniques:a social topology based approach and a tweet content based approach. And we studied latent factor model and several variant models in collaborative filtering. Based on the previous work, we developed a generalized latent factor model (GLFM) that makes use of heterogeneous data to recommend microblog friends. And we demonstrated how to simulate various latent factor models with GLFM. Using GLFM model, we just need to redefine input feature if we want to add new data source to the system. So we don’t need to redesign and implement algorithms. We developed feature extraction methods for common microblogging data sources such as social network, demography, tweet content, temporal dynamics etc. and used GLFM to model microblog friends recommendation. The experimental results showed that, GLFM can integrate multiple data source and obviouly improved the recommendation quality.