Research On Group Recommendation Based On Interaction Behavior Analysis

In recent years,with the development of social networks,group activities have become more and more common,which has promoted the rapid development of group recommendation systems.Different from traditional personalized recommendation,group recommendation aims to provide a group of users with services that meet their preferences.In the group recommendation system,there is a large amount of historical interaction information and potential interaction relationships among groups,users,and items.The processing of interactive information and the inference of interactive relationships belong to the category of interactive behavior analysis.These technologies have an important impact on mining user and group preferences.However,most of the current group discovery methods lack explicit coding of interactive information when embedding user features.So,these group discovery methods cannot fully reflect the potential similarity among users and lead to the poor performance of group discovery.In addition,most group recommendation methods cannot adequately model multiple interaction relationships when obtaining group preferences and they tend to ignore the dynamic changes of user preferences within the group.This type of group recommendation method ultimately achieves poor results.In response to the above problems,this thesis conducts research on group recommendation based on interactive behavior analysis.The main contributions are as follows:From the perspective of group discovery,this thesis proposes a group discovery method based on structured interactive information processing.The method first uses graph convolutional networks to explicitly encode a variety of structured interactive information which include user-item interaction graph and user social graph to obtain user feature representations.Then,the method adopts a group division model based on multi-task learning.This model can modify the user's feature representation and divide the group reasonably by training user preference prediction tasks and user clustering tasks jointly.Through simulation experiment comparison,it is concluded that the group discovery method proposed in this thesis improves the similarity of users within the group and the performance of recommendation.From the perspective of group recommendation,this thesis proposes a group recommendation method based on dual interactive relationships reasoning.Based on the realization of group discovery,the method first uses the GRU network to extract the user's dynamic preferences,and then comprehensively infers the interactive relationships which include user-to-user relationship and userto-item relationship to obtain the local influence weight and global influence weight of users within the group through the local and global attention units.Then,the method uses these two different influence weights to aggregate user preferences to extract group preferences.Finally,the recommendation is completed by calculating the similarity between group preferences and item features.The results of simulation experiments show that the group recommendation method proposed in this thesis improves the overall satisfaction of group users with recommended items.Based on the above methods and theories,this thesis constructs a group recommendation prototype system based on interactive behavior analysis,which is a movie group recommendation system based on Movie Lens data.Its construction includes requirements analysis,outline design,detailed design,specific implementation and other steps.Moreover,the group recommendation system designs multiple functions such as group discovery,group recommendation,personalized recommendation,movie ranking,and user evaluation feedback.The prototype system realizes the friendly combination of the theory and application scenarios in this thesis,and verifies the feasibility and effectiveness of the group discovery and group recommendation methods proposed in this thesis.