Deep Recommendation Approach Integrating Graph Structure Information

With the development of the Internet and the explosive growth of digital content,people are facing too many channels for information selection and acquisition.In order to deal with the problems of information overload and personalized needs,recommendation systems have become an essential module in all applications that provide services to users.They can quickly and accurately search for valuable information from massive data,and provide personalized recommendation content to users by analyzing their specific interests and behaviors.There are various types of data in recommendation systems,most of which can essentially establish topological associations and be constructed into graph structures.For example,there are user-item interaction graphs,social network graphs on the user side,knowledge graphs on the item side,and sequence graphs,all of which provide rich semantic information to help the system better understand the relations and features between users and items.Graph neural network in deep learning is widely used in recommendation systems due to its powerful ability to model graph structures and capture complex and implicit high-order collaborative signals between nodes.This thesis conducts in-depth research on deep recommendation methods that integrate graph structure information,present corresponding solutions for the unresolved problems and difficulties in previous research works.The main research contents of this article are summarized as follows:(1)This article focuses on the relation modeling problem in knowledge graph recommendation methods,and proposes a relation-fused graph attention network called RFAN,which improves the limitation of traditional graph attention networks that can only model node features,making it suitable for relation modeling in knowledge graphs.RFAN can encode high-order relation features while preserving the graph structure,and extend multi-layer information iterative propagation to update the representations of users and items.In addition,RFAN adopts an efficient non sampling strategy to train the entire dataset and jointly optimize knowledge graph embedding and recommendation tasks through multi task learning.Comparative experiments are conducted on CTR prediction and Top-K recommendation tasks with existing baseline models,demonstrating the excellent performance of RFAN in modeling knowledge graph relations.(2)This thesis focuses on the problem of noise triplets in knowledge graph recommendation,and proposes a collaborative attention network called DCRAN that combines entity differentiable sampling.During the entity sampling process,DCRAN applies a differentiable sampling strategy to reduce the number of noise triplets.In addition,DCRAN designs two different relation-based learning strategies for the user-item interaction and knowledge graph,and extracts the collaborative interaction signal from the interactions to customize the external knowledge,thereby distinguishing the different contributions of the same triplet to the user-item interaction.Similarly,experiments are conducted on CTR prediction and Top-K recommendation tasks,proving that DCRAN can alleviate the impact of noise knowledge on the model and further improve the performance of recommendation.(3)This article alleviates the problem of insufficient exploration of users’ potential semantic categories in social recommendation,proposes graph neural network-enhanced and cross-view semantic contrastive learning framework GSCL4 SR,which constructs the contrastive learning objectives from the structural and semantic perspectives respectively.From the structural perspective,we design a graph neural network enhanced graph contrastive learning module,using two graph learners to learn node representations and provide a comprehensive and rich feature representation for the model.From the semantic perspective,we design a cross-view semantic contrastive learning module that integrates user semantic information from interaction graph and social graph to narrow the semantic differences between the two views.A large number of experiments on three datasets of different scales show that GSCL4 SR can deeply mine cross-view semantic information of users and enhance users’ semantic representation.(4)This thesis focuses on the modeling of user dynamic interests in sequential recommendation,and proposes a multi-feature fused collaborative attention network called MASR.MASR not only incorporates temporal and positional features into node features,but also adaptively learns the importance of these two features,and then designs a collaborative attention network to clearly encode unified temporal signals and sequential patterns.In addition,MASR also considers the issue of node semantics and introduces a semantic enhanced contrastive learning strategy to enhance the semantic association between similar nodes.Extensive experiments on three publicly available datasets show that adaptive fusion of temporal and positional features can better model user dynamic interests,thereby improving the accuracy of recommendation.