Research On Personalized Recommendation Algorithm Based On Multi-Interest Of Users

With the advent of the information age,the volume of data such as news information,commodity information,entertainment video and so on show an amazing growth rate.How to extract the required information efficiently and conveniently from vast information has become a pressing issue that must be addressed.Personalized recommendation algorithm actively recommends the information that users are interested in by modeling users’ interests,which not only saves users’ time and energy,improves their quality of life,but also creates huge value for enterprises.Now,the study of personalized recommendation algorithms has yielded remarkable results,but with the uninterrupted improvement of the living standard of the public,some flaws are revealed.For example,users’ interests are diverse and constantly changing so that they cannot be accurately represented;Only considering users’ historical interest leads to the lack of timeliness and freshness of recommendation results;The lack of semantic relation mining affects the accuracy and generalization ability of the model.The main research contents of this thesis include:1.In view of the problem that personalized recommendation algorithm cannot accurately express user interest using a single vector,a multi-interest recommendation algorithm based on self-attention routing algorithm and Transformer is proposed.First,Transformer is used to mine the sequence relations in the user’s previous behavior data,and then the improved self-attention routing algorithm and capsule network are used to extract multiple interest vectors from the behavior vectors to express the user’s multiple interests,so that the multi-interest advice can be realized.The results of the experiments suggest that this algorithm can successfully increase the accuracy of user interest expression,as well as the precision and variety of recommendation results.2.In view of the problem that existing personalized recommendation algorithms only consider users’ historical interests,which leads to the lack of diversity and timeliness of recommendation results,a multi-interest recommendation algorithm combine users’ future interests is proposed.In this algorithm,part of the historical behaviors of neighboring users are taken as the potential behaviors of target users,and the future interests of target users are extracted by using the multiinterest extraction strategy in research content 1,and then the historical and future interests of users are combined to make the recommendation.Experiments have demonstrated that the proposed approach is both effective and practicable.3.In view of the problem that traditional personalized sequence recommendation algorithms mostly ignore semantic relations between items,resulting in ineffective use of global similarity between items and insufficient model generalization ability,this thesis introduces semantic relation and combines it with sequential relation through Graph Embedding technology.A highquality Graph Embedding is constructed by Light Graph Convolutional Network(Light GCN),which is used in the multi-interest recommendation algorithm proposed in research content 2.Experimental results suggest that Light GCN’s Graph Embedding technique,which combines sequential and semantic relation,increases the recommendation algorithm’s accuracy and generalization ability.