Research On Group Recommendation Method Based On Attention Mechanism

The recommendation system has developed extremely rapidly in recent years.The group recommendation system is based on a multi-person recommendation scenario and has emerged at the historic moment in response to the needs of group recommendation.Different from other single-person recommendation systems,in addition to recommending to a single user,the main task of a group recommendation system is to recommend for groups of multiple users.Usually in a group,due to the influence of group characteristics,social factors,etc.,the contribution rate of each member of the group to the group is different.Therefore,one of the most difficult issues in group recommendation research is how to integrate the group members' preferences,so as to meet the preferences of all members as much as possible.This thesis first proposes a self-attention group recommendation(SAGR)model based on self-attention mechanism and neural collaborative filtering(NCF)to model user interaction data and learning groups.Group potential preference representation.In view of the problem of incomplete consideration of group preferences caused by the attention network that can only obtain user item interactions,self-attention mechanism is used to mine interactions between users,so that the weight of each user in the group can be dynamically adjusted to solve the group preference fusion Problems to improve group recommendations.Secondly,this thesis also designs an Attentional Factorization Machine Group Recommendation(AFMGR)based group recommendation model.Because of its initiative to mine cross-cutting features and excellent generalization performance,factorization machines are widely used in project recommendation and CTR(Click-Through-Rate)estimation scenarios.However,it can only introduce second-order feature interactions,but cannot dig deep interaction information.Therefore,in order to fuse group preferences,the attention factor decomposition machine is combined with group recommendation.Each user in the group is regarded as a "feature".By capturing the importance of two users' interactions and thehigh-level interaction information between features,the weight value is improved,and ultimately the model prediction effect is better.Finally,this thesis compares the two methods with similar methods on the CAMRa2011 and Movielens datasets,and performs user rating prediction and Top-K recommendation.The experimental results show that compared with the current state-of-the-art algorithm AGREE,this method SAGR and AFMGR has more prominent performance.Among them,on the data set CAMRa2011,the accuracy indicators HR(Hit Ratio)and NDCG(Normalized Discounted Cumulative Gain)of SAGR have increased by about 1.6% and 1.3%,respectively,and the HR and NDCG of AFMGR have increased by approximately 1.9% and 1.5% respectively.On the dataset Movielens,the accuracy indicators HR and NDCG of SAGR have increased by approximately 0.9% and 2.3%,and the HR and NDCG of AFMGR have increased by approximately 1.7% and 4.5%,respectively.It can be proved that the method in this thesis can effectively improve the accuracy of group recommendation and the satisfaction of group users.