Research On Recommendation Algorithms Based On Network Representation Learning

With the development of Internet technology and mobile communication technology,the unprecedentedly developed information age has enriched and facilitated people's lives while also brought the problem of information overload.The recommendation system,as an effective method to solve the problem of excessive information,can help users filter information and provide accurate and targeted recommendation services.It has been widely used in e-commerce,social networking,audio-visual entertainment and other fields.In recent years,research on network representation learning technology has achieved outstanding achievements,and its ability to express network structure information provides new ideas for the research of recommendation algorithms.The thesis takes network representation learning and recommendation algorithms as the research content.The research purpose is to use network representation learning technology to improve the performance of recommendation algorithms and enhance the scalability of recommendation algorithms.The main work is as follows.For the recommendation problem in the homogeneous network,the thesis proposes an improved random walk based Attribute Network Embedding(RANE).This algorithm corrects the sampling results of random walk,and in the node vector learning part,the attribute information of the node is used.Specifically,in the random walk part,it is proposed to allocate the number of walks of each node according to the importance of the node,and control the walk with a certain probability to generate sequences of different lengths;in the representation learning part,based on the Skip Gram model,comprehensively considering the influence of the distance between the context node and the central node,a node vector learning model that can fuse node attribute information is proposed.Finally,it is applied to the recommendation task,and comparison experiments on multiple public data sets show that the algorithm can effectively improve the recommendation accuracy of the algorithm and achieve good results in the cold start recommendation problem.For the recommendation problem in the user-item bipartite network,the thesis proposes a Bipartite Network Convolutional Collaborative Filtering(Bi-NCCF)algorithm based on network representation learning.Specifically,the algorithm first decomposes the bipartite network into a homogenous network of users and items,and then uses the Graph SAGE algorithm in the two homogenous networks to obtain the user and item feature vectors that combine network space information and attribute information.The outer product operation is used to enrich the relevant representations of the user and item feature vectors in various dimensions.Finally,a convolutional neural network is designed to learn the complex interaction information between the user and the item,using the trained model predicts the interaction probability between the user and the item,and implements Top-N recommendation.Comparative experiments on multiple public data sets verify the feasibility and effectiveness of the algorithm.