Research On User Profiling Based On Topic Modeing

The development of social network brings about the explosion increasment of social data. It proposed the needed recommendation system. User profiling is one of key issues in recommendation system. In recent years, the trend of social network development is the diversity of data modalities and decrease of requirement of user interaction. The multi mode provided more extended possibility to user profiling.In this thesis we focus on the new social network – content curation social network, to study problem of user profiling, the details are as follows:1. We proposed a user profiling method by combining topic modeling and pointwise mutual information(TM-PMI). The idea of LDA model, the principle and the method of LDA inference algorithm--Gibbs sampling is a comprehensive description of the method of user text information modeling. Users personal topic words are extracted by combing the pointwise mutual information and topic models for users description, and a personalized user profile could be obtained. We design a user study method for the recommended evaluation. The experimental results show the TM-PMI expression ablitity is better than the traditional LDA model.2. We proposed a multi-level LDA(MLLDA) for modelling users' latent interests on content curation social networks. By analyzing the repin path in the content curation social networks, we use the descriptive texts for pins as well as the repin paths' ID of users for user profiling. We proposed the two dimensional LDA using both user description and repin paths. The user profiling results are used for community discovery. And we recommend users by calculate the Jensen-Shannon divergence between the current user and each user ID based on the proposed user profiling algorithm. The experimental results on Huaban showed that MLLDA is effective for community discovering and user recommendation.