Information Fusion Recommendation Based On Convolutional Graph And Neural Collaborative Filtering

The main goal of the Recommender System(RS)is to provide valuable and targeted information proactively based on user preferences.Collaborative filtering(CF)is a successful method among recommendation algorithms.However,the method-based collaborative filtering algorithms is generally limited in performance because of the problems of cold start and data sparsity.In order to solve these problems,utilizing auxiliary information such as content information or social networking information is a very promising approach.In recent years,many methods that make use of deep learning technologies have attempted to fuse auxiliary information into recommender system,which performance has improved when compared to the traditional methods.However,few recommendation models can handle graph structure information.Graph convolutional Networks(GCN),a graph convolutional network used in this paper,is able to fuse rating information with a variety of auxiliary information and then predict unobserved elements of rating matrix.Nevertheless,the previous graph convolution method of collaborative filtering is based on the traditional matrix factorization method,which only using the linear inner product to combine the embedding vectors and predict the rating.This method also leads to the reduction of the recommendation performance.Moreover,the existing nonlinear Neural Collaborative Filtering(NCF)cannot make good use of auxiliary information for rating prediction.The above problems have imposed limitation on the development of the recommendation system models.For the problems of the application of recommendation system based on graph convolutional network and nonlinear neural collaborative filtering,this paper make the following contributions: 1.This paper proposes an end-to-end information fusion recommendation model based on graph convolution and neural collaborative filtering.This model can not only take advantage of the users' rating information,but is also able to fuse the users' and items' respective auxiliary information to finish recommendation task.The input matrix is processed in an intuitive and efficient way,which facilitate the convolutional encoding calculation of the rating matrix and the auxiliary information.2.In order to overcome the shortcomings of traditional linear-based collaborative filtering model in the aspect of feature representation capability,this paper incorporate the matrix factorization method based on neural network to improve the graph convolutional recommendation process,and utilize the graph convolution model to fuse auxiliary information.The advantage is that the feature embedding output of the graph convolution model produces an improvement of the prediction result in the neural collaborative filtering process.3.The model and algorithm presented in this thesis base on the Tensor Flow deep learning framework and Python programming language.The experimental results show that the improved algorithm proposed in this paper is able to improve the performance of recommendation,when compared with the current popular recommendation methods on several real data sets.