Research On Social Recommendation Algorithm Of Graph Neural Network Based On Heterogeneous Information Network

Thanks to the rapid development of the Internet,people can easily obtain a large amount of information,but they also face the problem of information overload.In order to help users find items of interest in massive amounts of data,personalized recommendation system came into existence.However,in most real data sets,the interaction between users and recommended items is relatively rare,which leads to the model dose not have enough information to learn user preferences,resulting in cold start problem.To alleviate the cold start problem,existing research works have introduced auxiliary information in recommendation,such as item attribute characteristics,user social information,etc.These works have achieved good results,but the following shortcomings still exist: 1)The modeling method used is not perfect,and the internal information of different meta-path forms is not fully excavated;2)The information fusion method is monotonous and simple,which cannot aggregate rich contextual information inside the meta-path effectively for the target node.Both of these problems make user and project characterization incomplete.In response to the above problems,a heterogeneous graph network recommendation framework based on the attention mechanism is proposed.It gathers different levels of information based on the graph neural network,and aggregates information for the target node.By doing this,we can obtain richer user embedding and item embedding which have a closer relationship with the meta-path context content.Finally,experimental verification and performance evaluation are carried out on two real data sets with different characteristics.The contributions of this work are as follows:(1)Aiming at the problem of insufficient utilization of meta-path information,information of edge is added,and an information aggregation method of meta-path instances where user projects are located based on relational rotation is designed.While retaining the node vector information within the meta-path,the utilization of node location information is increased.Experiments show that compared with the baseline model MCRec,the accuracy,recall,and normalized loss cumulative gain are increased by 8.02%,5.05% and 3.22%,respectively.(2)Aiming at the coarse-grained problem of heterogeneous information mining process,an attention mechanism-based information fusion strategy of different meta-path instances where user items are located is proposed.Using a multi-head attention mechanism to gather meta-path instance information from different sources can overcome the instability caused by heterogeneous graphs.Experiments show that compared with the baseline model MCRec,the accuracy,recall and normalized loss cumulative gains are increased by 5.63%,2.72% and 2.96%,respectively.(3)Combining the above two points,a personalized recommendation experiment is carried out on two real data sets with different characteristics.The experiment shows that the accuracy,recall,and normalized loss cumulative gains on the Last FM data set with relatively even data distribution have reached 0.52,0.53 and 0.88,which are 9.50%,6.09% and 3.64% higher than the baseline model MCRec respectively.On the Movielens dataset where most user projects are sparsely distributed,the three indicators reach 0.41,0.26 and 0.79,respectively.Compared with the baseline Model MCRec increased by 20.30%,21.17% and 14.25% respectively.The cold start comparative experiment also shows that with the increase of data sparsity,the improvement effect of the algorithm is more obvious.Among them,the three indicators on the sparsest data set have increased by 24.94%,26.54% and 19.82% respectively.The work of this article will help e-commerce platforms recommend more products which users are interested in,which has practical significance to a certain extent.14 figures,13 tables,and 42 references are contained in the dissertation.