Utilizing Graph Structure In Recommender Systems

With the development of the Internet,people can obtain a large amount of various information online,which greatly enriches people's life.Too much information,however,also makes people take a lot of effort to find what they are really interested in.Recommender systems,which aims to alleviate the information overload problem,can provide personalized recommendation of items for users.Interactions between users and items in the recommendation scenario can form graph structures,which contains rich information to further improve the performance of recommender systems.Graph structures can boost the performance of recommender systems,especially in sequential recommendation and cross-domain recommendation.Sequential recommendation models the user's dynamic preferences within a short period of time,rather than only considering the user's longer-term static preferences.Besides the user and target item,the behavior sequence,consisting of items that the user has interacted with,is also used as input of sequential recommendation.The sequential dependence and the matching patterns can be learned by modeling the user's historical behavior sequence.Existing methods models the relation between the target item and items in sequence with sequence modelling modules,and then learns the matching patterns of these related items.However,none of these methods explicitly model the high-order collaborative similarity between the items.The high-order collaborative similarity can reflect the complex and potential relationship between items,and ignoring it will lead to the lack of matching patterns among items,which will limit the performance.Cross-domain recommendation aims to alleviate the data sparsity issue by leveraging the knowledge from an additional related domain.The interactions in a related source domain are used to improve the performance in the target domain,or achieve mutual improvement.Classical methods mainly base on mapping and transfer techniques,confronting with the issue of lacking in capturing high-order relations.Recently,a few methods introduce the graph neural networks to capture high-order information,and achieve superior performance.Items in cross domains,however,are totally different types of nodes,the heterogeneity leads to a noisy representation by aggregating heterogeneous neighbors directly.Moreover,user preferences in different domains are also different,treating cross-domain information samely with in-domain information during aggregation deteriorates the heterogeneity issue.The above issues in sequential recommendation and cross-domain recommendation are critical when utilizing graph structures in recommender systems.To address these issues,this thesis conducts related research,the main contributions include:For sequential recommendation,a Graph-Enhanced Sequential Recommendation model is proposed to address the issue of lacking high-order collaborative similarity modeling.Firstly,the item relation graph is constructed by connecting edges between the items which are in a certain size window based on a user's behavior sequence.The item relation graph contains the high-order collaborative similarity relation between the items within a short time range.The graph attention networks are used to aggregate the graph to capture the higher-order information,and then the item representation containing higher-order information is obtained.Then,the item representation with rich high-order information is fed into the gated recurrent unit and the attention mechanism for sequence modeling,thus more abundant item dependency and matching patterns between items can be learned.For cross domain recommendation,a Metapath-guided Cross Attention Network is proposed to address the issue of the heterogeneity between nodes and domains when capturing high-order information.Firstly,the interactions of both target domain and source domain can construct a joint interaction graph,which is a heterogeneous graph consisting of three types of nodes.Then,the defined metapaths are used to aggregate the neighbors to get the node representation containing highorder information,which is not affected by the heterogeneity of nodes.The defined metapaths include in-domain metapaths and cross-domain metapaths.The model learns an augmented representation for a node by aggregating with in-domain meta-paths,and then use it as a query to attentively extract information with cross-domain metapaths.In this way,noisy information from cross-domain can be further filtered.Moreover,in cross-domain metapath aggregation,additional transformation of neighbor representation further alleviates the heterogeneity between domains.The thesis focuses on utilizing graph structures in recommender system,and proposes two improved models respectively for sequential recommendation and crossdomain recommendation.Extensive experiments are conducted and further analysis verifies the effectiveness of the proposed models.