Reserch On Key Technologies In Social Media Oriented Content Recommendation

Web 2.0 technologies boom the social media and make the user to be the content generator at the same time. And social media is becoming a powerful tool for users to get information and participate in social activities. Social media is also affecting our society, politics and economy etc. But social media makes the information overload, explosion of the content and fragmented spread.As the large scale, diversity, timeliness and power law characteristics of social media content, users are very easy overwhelmed by information overload problem.Collaborative filtering is the most widely technology in the traditional recommendation systems. And it uses the users’history ratings on the items to mining the users’preference and predicts the ratings. But it confronts problems including data sparsity, cold-start etc. In this paper, we investigate into four different use cases in social recommendation to utilize the current technologies with social and context information to solve the data sparity and cold-start problems. And we also use our algorithms in the real book recommendation in the following four practical cases:just with user-item score, just with user behaviors, with uer-item score and user social information, with user-item score and context information.And we research on the corresponding cases based on community detection, combination of different prediction methods for recommendation, combination of community detection and the traditional collaborative filtering technology and recommendation based on the context information respectively.The major contribution of this thesis is as following:1) To overcome the low performance of the traditional algorithms casued by noise in social media, we propose two different combination methods to combine the traditional ones. The first one is modeling combination which combines different algorithms in modeling and the other one is prediction combination which combines the different algorithms in prediction. The modeling combination method considers both the local similarity of neighbor-based algorithms and the global similarity of model-based algorithms to modify the minimized function, while the prediction combination method uses the supervised learning algorithms such as LR, BLR and NN to minimize the error.The results on the dataset of library indicates the efficency of the two combination methods. And the best is RSVD-3+RSVD2-3(NN) which error is only 11 days in predicting the borrwing time.2) To utilize the prefrecne of groups in social media recommendation, we propose a recommendation algorithm based on community detection. The community detection method aims to expand the community kernels whic is composed of disconnected nodes to detect communities. The community detection algorithm is called LDK which finds communities with affective kernels. The recommendation algorithm based on LDK is proposed as RoL which considers the change of popularity with time. The results on the dataset of the library show that RoL(3) performs 6.58% better than Heats in HR and 5.69% better in ARHR.3) To overcome the popularity items’affect and the data sparity problem, we use the social information to incorporate in prediction and recommendation. And then we propose the memory-based collaborative filtering algorithm based on community detection NCFC and the model-based collaborative filtering algorithm based on community detection SCR. To detect communities in the weighted directed networks, we transformed LDK by modifying how to measure the effect of nodes and the fitness function.NCFC delete the neighbors with low similarity and use the global similarity to overcome the sparsity problem. NCFC is proper in both click recommendation and rating recommendation. And NCFC performs 6.82% better than RoL(3) in HR and 17.43% better in ARHR.In click recommendation,in rating recommendation, NCFC performs better than ICF in both MAE and RMSE. SCR use the neighbor features to add regularization and it performs 5.83% better than ASR in MAE and 6.01% better in RMSE..4) In order to improve the prediction accuracy, we fully consider the context information and incorporate it into the traditional collaborative filtering technology and propose the neighbor-based collaborative filtering technologies PLHS and CPHS and the model-based collaborative filtering technology SLUC. PLHS and CPHS combine the traditional similarity computation based on ratings with context information. And SLUC extends the existed method SLIM by adding the users’information in the regularization. The experiments on the real world dataset invalidate the efficiency of the algorithms. CPHS is better the existed RPBC 5.4% better in HR and 3% better in ARHR. SLUC performs better than SLIM in both HR and ARHR on dataset ML100K and the library dataset.