Research On Preference Aggregation Strategy Of Group Recommendation

The development of all kinds of smart mobile applications makes it difficult for users to obtain information that accurately matches them from massive amounts of data.Recommendation system can provide users with personalized services intelligently and quickly,effectively solving the above information overload problem.However,whether in real scenes or virtual communities,users tend to participate in activities in the form of groups,for example,watching a movie as a family or traveling with friends.Therefore,providing an accurate recommendation service for a group of users,that is,group recommendation,has become a new service model.Group recommendation not only needs to take into account the differences in the preferences of each user in the group,but also considers that group decision-making is affected by the interaction between users in the group.Therefore,how to use the available data to integrate the preferences of group members to build a group model is the group recommendation system critical mission.The commonly used preference aggregation strategy uses static methods such as mean strategy to fuse group member preferences,failing to fully consider the different contributions of group members' individual characteristics to group decision making,ignoring that the user's ideas are influenced by other users in the group,which in turn affects group recommendation performance.In summary,this article proposes two novel preference aggregation strategies,the main contributions are summarized as follows:(1)In this paper,a preference aggregation method based on user importance is proposed,which learns group preference representation by modeling user interaction data.Because users have different importance in group decision-making,we used attention mechanism to model the higher-order characteristics of group members,that is,the importance of each user.These dynamic user weights are aggregated to form group representation vectors.In addition,users' decisions in group decision-making will be affected by other users in the same group,so a neural network that combines the multi-head attention is designed.It automatically learns the high-order interaction between users to obtain the interdependence between group members,and finally make the model has the best prediction effect.(2)In order to further improve the satisfaction of group members with the recommendation results,a preference aggregation method embedded in the user's social attributes is designed.This method uses a dual-layer attention network to model group user preferences under the framework of representation learning.First,the user's social attributes are introduced to form a user preference representation vector through attention network,and then it is input into another attention network aggregation to form a group representation vector.Finally,through the interaction data between the modeling group and the item,the group's predicted value of the candidates is obtained.This paper evaluated the proposed methods on real datasets,and compared with the experimental results of the baseline methods.The experimental results show that the proposed method has better recommendation performance.