Research on Graph Network Recommendation Algorithm Based on Random Walk and Convolutional Neural Network

Abstract

As a general form of describing and modeling complex systems, networks widely exist in different scenes and tasks in various fields of the real world. Therefore, how to effectively calculate (graph Computing) and analyze (graph mining) data based on network structure has always been the core basic research direction in the field of computing science and data mining, and has been continuously studied by scholars from computer, sociology, biology and other disciplines. Network representation learning can better analyze the information hidden in complex networks, and with the help of graph neural network, it provides a universal method to solve various practical problems under the background of network structure data, which has attracted the common attention of academia and industry. At the same time, traditional recommendation algorithms generally analyze the user's rating data on items and then make preference recommendations. These methods have some problems, such as difficulty in extracting deep features, single data processing method, etc. These problems will lead to low prediction accuracy and unreasonable results. Therefore, in this paper. The recommendation algorithm for heterogeneous networks based on feature embedding is improved (RW-CNN), and the random walk algorithm of convolution and graph networks in deep learning is used to process the text feature of item names. Firstly, the network topology similarity calculation module is used to measure the consistency and complementarity of each view network, and then the word vector learning technology is used to generate the representation vector of multi view network: the word vector module is used to distinguish the complementarity between views, and the hierarchical hidden state based on multi views is used to extract these information, To retain the unique complementary information of each view; The word vector module is used to recommend multiple views with stronger consistency. Through the middle hidden state shared by multiple views, the graph convolution technology is used to aggregate information and realize the fusion of consistent information. Finally, the learned representation vector is used for link prediction and node classification tasks. The experimental results on a variety of real data sets show that the effect of this model is significantly improved.