Research On Relational Graph-based User Behavior Sequence Modeling

For a long time,sequential recommendation technology has been the hot topic of recommender system research.Most sequential recommendation models focus on learning from sequences of items that users interact with in a purely data-driven manner and fail to effectively utilize information about users,items,and the relations between users and items.Due to the large amount of interactive data and the improvement of computing platforms,models based on graph neural network technology have achieved great success in sequential modeling tasks in recent years.However,there are still some defects in the current research of sequential recommendation technology based on graph neural network:(1)the neglect of prior knowledge of relations between items;(2)the neglect of temporal dynamics of user and item interactions.Based on the above problems,this paper conducts in-depth research on two types of sequential recommendation tasks respectively,i.e.,modeling the relations between items using knowledge-driven relational graph and encoding user-item sequential interaction using data-driven relational graph,and designs corresponding user behavior sequence modeling model.The main work and contributions of this paper can be summarized as follows:For the problem of the neglect of prior knowledge of relationships between objects,this paper focuses on learning relations between items to improve the sequential recommendation task.Specifically,this paper focuses on the critical substitutable and complementary relations between items in e-commerce platforms with the aim of leveraging these explicit item-item relations to facilitate sequential item recommendations.There is currently little work that considers the relations between items in sequential recommendation task.However,they both ignore high-order relation modeling.Therefore,this paper proposes a graph based sequential recommendation model SCG-SPRe which considers characterizing substitutable and complementary relations.This method proposes to model the high-order relation between items through a interactive graph neural network(knowledge-driven relational graph),and use the kernel-enhanced Transformer model to characterize the temporal dynamics of user behavior sequence.The experiment proves that the effect of SCG-SPRe model exceeds all the current baseline models,and proves the necessity and effectiveness of considering the substitutable and complementary relations.This work has been published in TOIS-2021.For the problem of the neglect of temporal dynamics of user and item interactions,this paper focuses on learning the dual dynamic representations of users and items to generate recommendations.Temporal dynamic modeling is critical for sequential recommendation,but most existing studies only focus on the user side and ignore the sequence paradigm of unexplored item side.Although some studies have examined the dual dynamics of user-item relationships,complex user-item interactions are not fully utilized from a global perspective.Therefore,this paper proposes a dual dynamic representation learning model DRL-SRe based on dynamic graph neural network.In this method,a time-sliced graph neural network(data-driven relational graph)is proposed to model the global interaction information of user and item,and auxiliary temporal prediction task is proposed to capture the fine-grained time information in user behavior sequence.Experiments show that the DRL-SRe model outperforms all the current baseline models,and proves the great potential of dual dynamic representation in sequential modeling tasks.This work has been published in CIKM-2021.In summary,from the perspective of task characteristics,this paper researches and explores the application of different relational graph neural networks in sequential modeling scenarios.Experimental results on several real-world and public data sets demonstrate the effectiveness of the proposed methods.