Research On Knowledge Graph Heterogeneous Information Collaboration Based Recommendation Methods

With the rapid development of the Internet and the explosive growth of big data,people are facing serious information overload when obtaining products and services.How to fully exploit the potential value in massive data using artificial intelligence and big data-related technologies,and provide efficient and high-quality services to consumers,has become an urgent problem to be solved.Recommender systems,which analyze users’ historical behaviors and interests to provide personalized recommendations,are playing an increasingly important role in the fields such as e-commerce,social networks,and news media.In recent years,thanks to the prosperity of deep neural network algorithms,research on deep learning-based recommender systems has made significant progress.Traditional recommendation methods mainly rely on users’ historical behaviors and are susceptible to issues such as data sparsity.To alleviate these problems,knowledge graphs(KGs)-based recommendation has attracted great attention from researchers.As a structured knowledge representation method,KG integrates different types of heterogeneous information such as entity semanticity,attribute hierarchy,relation relevance,topological complexity into a unified framework.KG heterogeneous information collaboration(HIC)of recommender system refers to the collaborative processing of heterogeneous features in different information carriers,thus improving recommendation performance.However,existing methods have the following problems when representing and modeling KG heterogeneous information:(1)For entity HIC,existing methods usually adopt more coarse-grained neighbor propagation and entity aggregation methods,thereby ignoring the enhancement of the personalized preferences hidden in heterogeneous entity groups;(2)For attribute HIC,existing methods lack exploration of multi-level heterogeneous hierarchical information hidden in item and entity attributes,and cannot accurately capture users’ fine-grained preferences;(3)For relation HIC,existing methods ignore the relevance information of potential relations in KG and cannot perceive the impact of different relationship contexts on information propagation;(4)For topological structure HIC,most methods rely on a single type of Euclidean embedding space,which cannot provide sufficient learning capacity to capture complex topological structure information in KG.Besides existing recommendation methods usually ignore the interaction between different spatial features.For the above problems,with the research focus on the HIC in knowledge graph,and with the research goal of improving the KG-based recommendation performance,this thesis has systematically conducted the research on the knowledge graph heterogeneous information collaboration-based recommendation methods,involving the different perspectives such as entity,attribute,relation,and topology.The research contents are as follows:(1)For the research of how to capture heterogeneous semantic information in KG entities and achieve fine-grained collaborative aggregation of neighboring entities.This thesis proposes a graph attention network-based relation embedding model(GRE).This method innovatively clusters user interaction item and its adjacent entities into triple set.Then triple set is further subdivided into triplet groups based on the relation type between the head and tail entities.This operation ensures that the head entities and relations in each triplet group are the same while the adjacent entities are different.Subsequently,the graph attention network algorithm is used to aggregate the tail entity information into the head entity for the triple group.Clustering and aggregating the entities in KG can achieve refined collaborative aggregation of neighboring entities,thus accurately capture the fine-grained user preference.(2)For the research of how to model heterogeneous hierarchy information in KG attribute and achieve hierarchical collaborative representation of triple knowledge.This thesis proposes a triple multistage clustering-based hierarchical attention model(MCHA).This method innovatively clusters items and entities into item clusters and entity clusters based on the attributes of user interaction items and the relations of the adjacent entities.This operation is called multistage clustering.Subsequently,a hierarchical attention network is constructed by stacking multi-layer attention mechanisms.The attention coefficient is used to measure the contribution of multi-level heterogeneous information to capturing user interaction intentions.In addition,this method explicitly embeds relation information into entity clusters,thus further enhancing the accuracy of hierarchical collaborative representation of knowledge.(3)For the research of how to mine heterogeneous relevance information in KG relations and achieve differentiated collaborative perception of entity relations.This thesis proposes a relation-aware attentional graph convolutional network model(RAAGCN).This method innovatively replaces the binary adjacency matrix in graph convolutional networks with the attention weight adjacency matrix.RAAGCN can achieve adaptive aggregation of entity pairs with explicit and implicit relations.Meanwhile,considering the personalization of user preference,a translation model is introduced into the adjacency matrix construction process.This operation aims at obtaining the plausibility coefficients between head and tail entities under specific relationships and capturing contextual information of the relations in KG.In addition,based on RAAGCN,recurrent neural network as well as gating mechanism,this thesis innovatively refines user preference into interest preference and rating preference.It can improve the differential preference perception ability of recommender system.(4)For the research of how to explore heterogeneous diversity information in KG topological structures and achieve personalized collaborative interaction of structural patterns.This thesis proposes a mixed-curvature manifolds interaction learning model(CurvRec).This method migrates graph convolutional networks from traditional Euclidean space to nonEuclidean mixed-curvature manifold space to capture complex topological structure information in KG.At the same time,the thesis also designs the Ricci curvature aware neighbor aggregation method to capture fine-grained local structural features.Different manifold spaces focus on capturing different topological structural patterns.In order to enhance each other’s spatial features,interaction learning shcema is introduced to achieve personalized collaborative interaction of topological structural patterns.In summary,this thesis studies the problems of heterogeneous information collaboration of knowledge graph-based recommendation.The corresponding HIC recommendation methods are proposed,specifically involving fine-grained collaborative aggregation of entities,hierarchical collaborative representation of attributes,differentiated collaborative perception of relations,and personalized collaborative interaction of structures.This thesis comprehensively and systematically elaborates on the enhancement of HIC on the performance of KG-based recommendation.Based on multiple public datasets,this thesis has conducted sufficient experiments to verify the effectiveness of the proposed methods in Click-Through-Rate prediction and Top-K recommendation tasks.