Group Discovery And Recommendation For Social Networks

With the rapid development of information technology,the amount of data on the Internet grows exponentially.In order to obtain effective information from massive data,recommendation systems which can meet the demands of different users arise.Recommendation systems for social networks are used to promote the development of services,as a consequence to study how to aggregate social information more effectively into recommendation system become a key issue.However,most current group discovery methods for social networks ignore the combination of explicit and implicit user preference information,which results in unreasonable group division.Moreover,most of the social network recommendation systems only consider the historical preference of users,and they ignore the issue of time migration of user preferences,it has a negative impact on the recommended results.In order to solve the above problems,this thesis concentrates on group discovery and recommendation for social networks.The main work is as follows:Firstly,a group discovery method based on multi-view learning is proposed to solve the problem of ignoring the combination of explicit and implicit preference information of users in most current group discovery methods for social networks from the perspective of group discovery.The method firstly extracts explicit multi-dimensional user preference information based on user history activity information,and the dynamic topic model is used to generate and update the user preference dynamically.Then,the user explicit preference information is fused with multi-view.In order to obtain more user information,implicit preference information is obtained by unsupervised learning and training.Finally,the user similarity matrix is used to divide groups.Through the comparison of simulation experiments,it is concluded that this method improves the similarity and recommendation accuracy of users within the group and reduces the recommended error rate.Secondly,since most existing group recommendation methods for social networks ignore the time migration of users preference,a group recommendation method based on deep learning is proposed to solve the problem of the time migration of users preference for social networks from the perspective of group recommendation.This method firstly mines users' historical preference information and conducts hierarchical clustering based on the topic content,then obtains the topic distribution of users' preference based on the LDA topic model,and dynamically obtains users' preference by adjusting the weight of time function.It extracts the characteristics of the user and the recommendation service based on deep semantic network.Considering user's social relationship will affect their service choice,this thesis modeled the communication process between users and converted it into a deep neural network to solve the problem and obtain the user service score.After that,the group recommendation is completed by the service scores of the group members.Through the comparison of simulation experiments,it is concluded that this method alleviates the static limit of user preference and improves the accuracy of the recommendation.Finally,this thesis designs a prototype system of group recommendation for social network and a demonstration of group service recommendation application simulating MicroBlog based on the above methods and theories.Moreover,this thesis introduces the components,operation flow and system structure of the system in detail.And it verifies the feasibility of the method and theory proposed.It shows the recommendation effect of the group discovery method based on multi-view learning for social networks and the group recommendation method based on deep learning for social networks in real social network.The prototype system realizes the friendly combination of the method theory and the practical application scenarios in this thesis,and demonstrates its effectiveness and practicability through the effect display.