Research On Collaborative Filtering-Based Method For Personalized Community Recommendation

In online virtual community, users can create groups or communities according to different topics, and also can join groups or communities created by others to discuss with each other, exchange information and so on. Today, the number of virtual communities on the Internet increases dramaticly. It is more and more difficult for users to find communities which they interested in from such a huge network. Therefore, community recommendation for users becomes a growing concern in the research.Existing community recommendation algorithms are vulnerable to overfitting and computational intensive problem caused by the huge quantity of data. Besides, these methods ignore the strength of the relationship between user and community. They also failed to consider the changes of user interest over time. When new users join some groups, they are not able to quickly update the model. In this situation, this paper studies how to relief these existing community recommendation problems. The main research work and innovations of this thesis are as follows:1. Proposed a soft-constraint based LDA community recommendation algorithm S-LDA, which select the number of user's posts on the community to measure the strength of the relationship between user and community. This algorithm considers each user as a document, the community that the user joined in as a word in the document, and the strength of the relationship between user and community as the number of occurrences of the word in the user document. Then the model parameters are inference by Gibbs sampling. Experimental results show the feasibility and performance advantages of the algorithm.2. Proposed an online update system framework to deal with the scalability problem of S-LDA. When a new user is added, maintain the original trained model parameters unchanged, and train a separate model for new user document. In this case, only a small number of iterations are needed to reach convergence, thus the computational complexity can get a greatly reducing.3. Proposed a time information based community recommendation algorithm, which take the time information into the modeling of user interest model. The impact of each user post behavior to the model is time-related decay. This effect is modeled with kernel density estimation method. The impact factor of the time information on user interest modeling can be used to weighting user-topic distribution. Experimental results show that the algorithm can enhance the performance of S-LDA.4. SO-LDA algorithm is implemented on LISER platform. In practice, the algorithm has good performance.