A Group Recommendation Algorithm With Consideration Of Social Relationship And Diversity Factor

Because of the rapid increasing of online information, it has been an urgent issuethat how to help users find useful information from a large amount of resource.Recommendation system is one of the important solutions to this problem, whichprovides personalized service to users efficiently and helps users find items or contentthey might be interested in. It has been researched and applied widely for itssignificance, and the recommending technology has been developed maturely.However, most existing recommendation systems cater to individual demands, whilewe need to recommend to groups in many situations, for example, recommendingrestaurants or tourist attractions to a group of friends and recommending TV programsto families. In addition, the popularity of Social Network brings about a large numberof online groups, generating the need to recommend to these groups as well.Facing to group users, we built a new group recommendation algorithm based onSocial Network Analysis (SNA), which combines advantages of existing algorithms.We mainly developed and improved the process of building group models by takinginto account of social relationship among users, different status of group members andthe diversity of opinions in groups. Firstly, we utilized SNA to calculate “distance”among users with consideration of social relationship, based on which, we searchedneighbors for the target user. Then we predicted ratings for the target user accordingto ratings of its neighbors. Thus, we built the individual rating model. Secondly, weused Respect Strategy to aggregate individual models into group models, necessaryweights in the process were obtained from “Relative Centrality” in SNA. Then weamended the group model on account of Diversity Factor and got the final groupmodel. Recommendations were made based on the group models.We did experiment to our method as well as comparative methods on the datasetoffered by Baidu. Experiment results showed that our method achieved higherrecommending accuracy.