Research On Recommendation Method Based On Multi-Interest Framework

Recommender system is an emerging and important research field,which refers to learning user preferences based on user interaction or scoring of items,to filter information or recommend items for users.At present,the recommender system has been widely used in practical scenarios such as ecommerce recommendation,information flow display,and personalized information.Different recommendation methods are applicable to different recommendation scenarios.In scenarios with certain historical interaction data,graph recommendation and sequential recommendation methods can be used to provide users with better and more personalized results.In scenarios where the data is extremely sparse,the cross-domain recommendation method can be used to explore the information contained in the auxiliary domain to alleviate the cold start problem of the current domain.Nowadays,a series of multi-interest sequential recommendation methods have emerged,which consider modeling user behavior sequences into multiple interest vectors,aiming to predict items that users are more interested in based on these vectors.However,such multiple interest sequence methods only learn multiple interests of each user from their recent interaction sequences,without utilizing the global context information contained in all users’ historical interactions.At the same time,current cross domain recommendation methods based on preference mapping also ignore the multi-interest preferences contained in user interaction sequences in the source domain.Therefore,this thesis proposes two new recommendation methods aimed at solving the above two problems respectively:(1)This thesis proposes a method named "GIMIRec: Global Interaction-aware Multi-Interest Framework for Sequential Recommendation".This method proposes the global context extraction module for the first time and designs a simplified graph neural network to learn items’ co-occurrence information from the global context.Specifically,this method first extracts the global context information from the interaction of all users in the training set according to some rules and obtain the global context embeddings of the items.Secondly,the time interval embeddings of items are then obtained from each user’s recent item sequence and time sequence.Thirdly,using the aggregation module,the above two embeddings are used to represent and learn the user’s recent item sequence.The user’s personalized embedding is obtained after the above steps.Finally,multi-interest representations of users are extracted from the user’s personalized embedding and used for recommendation prediction.(2)This thesis proposes another method,named "Cross-domain Recommendation Based on Multi-Interest Modeling".This method introduces multi-interest modeling into cross-domain recommendation tasks for the first time and makes full use of the user’s interaction sequence information in the source domain.Specifically,this method pretrains the scoring data in the source domain and the target domain to obtain the global embedded representation of users and items in each domain.Then,the self-attention mechanism is used to model the recent interaction sequence of users in the source domain.Subsequently,the multi-interest framework is used to learn multiple interests of users from the sequence.Then,the most relevant interests with the global embedding of the user are selected,and a personalized preference mapping bridge is trained for each user.Finally,the output of the preference mapping bridge is used as the user’s preference representation in the target domain to achieve the score prediction recommendation task.The experimental results show that both methods proposed in this thesis achieve the best performance among similar models without introducing external data.In the future,for the non-cold start scenarios,we can consider using the self-supervised learning method to filter items that are not of interest in the user’s sequence to improve the recommendation performance.For the cold start scenarios,it is possible to consider introducing auxiliary information from each domain and designing fine-grained interest mapping bridges to capture user’s preferences in the source domain,in order to achieve better cross domain recommendation results.