Research On Group Recommendation Based On Attention Mechanism

Recommender systems can effectively mitigate the problem of information overload and help users select items or services that they are interested in more quickly.As social animals,participating in group activities is an indispensable part of people's daily life.Those group activities include,for example,having lunch with colleagues and traveling with friends.Therefore,recommending satisfactory items for a group of users,that is,group recommendation,has become an important task in many information systems.The difficulty of the group recommendation system is how to aggregate personal preferences of group members to infer group decisions.Prior research on group recommendation has been focused on exploring various static strategies,that is,through some predefined strategies(e.g.,average,least misery,etc.)to aggregate the preferences of different group members.However,these static aggregation strategies are relatively simple and only aggregates members' preferences without considering interactions among members in a group,which is difficult to model complex group decisionmaking processes,resulting in suboptimal group recommendation results.Therefore,the ideal group recommendation should be able to dynamically aggregate members' preference to more accurately model the group's preference.In view of the above analysis,this thesis proposes a model combining fusion gating mechanism and multi-layer attention mechanism for group recommendation(Gated Attentive Neural Network,GANN).The main content of the thesis as follows:(1)This thesis investigates related technologies involved in the research work,such as group recommendation,collaborative filtering,and deep learning,and summarizes the research status of these related technologies.One this basis,this thesis elaborates the theory of related technologies that are closely related to this research work.(2)The GANN model proposed in this thesis first models group's preference through two parallel modules: an Item-based Aggregation Module(IAM)and a User-based Aggregation Module(UAM).Specifically,the IAM considers different users' preferences for different items and uses the attention mechanism to learn the weights of group members for different items,thereby merging to form a specific preference vector of the user group for different items,that is,the group representation vector aggregated by items' property.In addition,in order to better learn the interaction between users in a group and accurately model group preferences,the UAM uses a self-attention mechanism to learn the relationships between members within a group to form a group representation vector aggregated by members' relationship.Then the two learned vectors are selectively combined into a final representation of the group through the fusion gate.Finally,the model calculates a nonlinear matching function value between the final representation of the group and the candidate item representation to obtain a recommendation score for each candidate item.(3)To demonstrate the effectiveness of the proposed model,this thesis compares the GANN model with the current mainstream models on three real data sets in different fields.The experimental results show that for the four commonly used performance evaluation indicators of the recommendation system,compared with other baseline models,the GANN model has improved performance.