Research On Recommendation Algorithm Based On Heterogeneous Graph Convolutional Network

In the information age,people can access a lot of information everyday,but correspondingly,it also brings great challenges for people to choose the information they are interested in.With the explosion of information,recommendation algorithm has been widely concerned by academia and industry.There are two cores of the modern recommendation system: one is to learn the embeddings of users and items(that is,to convert users and items into vectorization representations),and the other is to model interactions,which are based on embeddings and reconstructing historical interactions.As one of the cores of modern recommendation systems,it is very meaningful to study methods for better learning the embedded representations of users and items.In order to better learn the latent feature representations of users and items,a model that combines heterogeneous graph convolutional network and multi-layer perceptron has been designed and implemented.On the basis of user-item bipartite graph,a new heterogeneous graph combining the user's reviews and the item's descriptions has been constructed,and the graph convolution network(GCN)has been applied to the heterogeneous graph.The heterogeneous graph convolutional network can make use of the auxiliary information such as reviews and descriptions and the structural information of the graph,so as to better obtain the latent representation of the user and the item.In addition,this paper uses the word vector that is pre-trained by Glo Ve to initialize review text or item's description information.Different from other recommendation models based on graph convolution network,which use GCN to get the latent embedded representations of users and items and use inner product as interaction function directly,Text NGCF combines multilayer perceptron with GCN,and uses MLP to continue the representation learning of users and items.In order to verify the effectiveness of the recommendation algorithm based on the heterogeneous graph convolution network,Text NGCF has been verified on Amazon's four public data sets,and compared with some current mainstream recommendation algorithms.In addition to the user-item rating data,the data set also uses reviews and item's descriptions as auxiliary information.The results show that the algorithm,which combines the heterogeneous graph convolutional network and multi-layer perceptron,performs better than other algorithms on the same data set,and can learn the latent feature representation of users and items well.