Research On Recommendation Method Based On Heterogeneous Graph Neural Network

In the current information explosion era,the generation of massive information has also led to the rapid development of technologies such as the Internet and big data analytics.While we are enjoying the convenience of modern Internet,we are also facing the problem of information overload.It is difficult for people to easily and quickly extract the required information from the huge amount of data effectively.The research on recommendation systems is geared to this realistic and urgent task,and is rapidly becoming a hot research topic in the field of data mining.Recommender system approaches can improve user experience by modeling data information and analyzing potential associations between users and items,thus mining user behavior preferences.Nowadays,recommendation systems have been widely implemented and practiced in many scenarios including music,shopping,books,movies,etc.,and have achieved great success in many commercial applications such as Amazon,Facebook and Taobao.Since traditional recommendation systems mainly consider the historical behavioral interaction between users and items,but the limited information interaction between users and items in practical application scenarios often directly affects the results of recommendations.And the sparse information of user behavior history also brings problems such as recommendation cold start.In recent years,many studies have integrated auxiliary information of users or items into the modeling of large-scale or scarce networks,forming a heterogeneous information network containing various information.Most algorithms are based on meta-paths to obtain similar users and learn to extract valid information through network representations.Although the fusion effect has been improved,most of them have the following problems:Due to the heterogeneity of information,different types of information features cannot be directly fused;The existing studies only model the structural association features or contextual semantic information of heterogeneous information,but not fusion modeling;Their recommendations are not credible when the data are sparse and contain noise.To address the problems that have existed,we proposed a heterogeneous graph neural network-based recommendation method which combines the theory related to heterogeneous graph embedding to model large-scale networks.The main work is as follows:(1)We present the multi-level attention mechanism of heterogeneous graph neural network recommendation model to address the problem that heterogeneous information cannot be directly fused.The heterogeneous information network structure is granularized into multiple independent coarse-grained networks by constructing different meta-paths.And then the graph attention mechanism and path-level attention mechanism are used to learn the potential features of users and items respectively.Finally,the fused fine-grained network recommendations are given.The experimental results show that the method can effectively improve the recommendation performance in both cross-sectional and longitudinal evaluations on realistic large-scale datasets.(2)In location-based social networks(LBSNs),most of the current research models the semantic context,ignoring the influence of the associative features of the structure on the model.Therefore,we present the multi-granular attention mechanism for point of interest(POI)recommendation model.From the different effects of attribute node embedding and structural meta-path embedding,the location similarity between POIs is computed by combining Triangular kernel function to achieve geolocation-based POI recommendation.The effectiveness of the method for POI recommendation is also verified on real data sets.(3)We present a multi-view heterogeneous graph neural network of POI recommendation model for location-based social networks(LBSNs)with sparse and noise-laden data.The heterogeneous graphs are subjected to coarse operations as node dropout and fine-grained link prediction operations by data enhancement techniques to generate multiple views for the purpose of graph structure perturbation and node enhancement,respectively.Then,the representation of interacting sparse nodes is improved by performing random walks on local subgraphs.Aggregate heterogeneous node features with hierarchical graph neural networks.The features of groups are obtained from a sequence of random walking,and the feature representation of user nodes is obtained by content embedding from different groups to different views for POI recommendation.The performance of the algorithm is evaluated on a real dataset and the experiments show that the model proposed in this paper has perfect performance.