Research On Personalized Recommendation Algorithm Based On Graph Convolution Networks

In the information age,although people can share resources and broaden their horizons,the massive amount of information will make users spend more time seeking for information that meets their requirement.It is precisely in order to solve the problem that users are difficult to choose,that personalized recommendation has received widespread attention.Personalized recommendation push information that user are interested in to user by analyzing user behaviors and marking the user's personalized preferences.At present,the main problem of personalized recommendation is sparse data.Exploring the user's interest pattern with assistance of additional auxiliary information of users and items is one of the effective solutions.In recent years,Graph Convolutional Networks have shown strong learning ability on graph data and have been widely applied in recommendation systems.This thesis mainly utilizes graph convolutional networks to analyze auxiliary information to improve the accuracy of personalized recommendation.The main research contents of this thesis are as follows:1.To accurately explore user preference patterns,this thesis proposes a graph convolution matrix completion via relation reconstruction(RE-GCMC)method.Firstly,the similarity measurement function is used to construct the user-user and item-item similarity graphs according to the user information,item attributes and historical interaction data,and the attention graph convolutional network is exploited to mine the nonlinear relationship between users and items from various graphs and capture the higher-order semantic information of different neighborhoods to learn latent feature representations of users/items with structural and content information.Secondly,multi-layer perceptrons fuse feature information and explore high-order feature interactions between users and items to improve the feature representation of users and items,thus predicting users' ratings of items that have not been interacted with.Finally,experiments on the recommendation datasets show that the proposed method is superior to baseline methods.2.In order to solve the problem of mining the high-order interaction between users and items from review information,this thesis proposes joint rating and reviews with graph convolutional networks(RRGCN)method.Firstly,reviews are mapped into word vector space by word embedding technology,and then CNNS is used to learn the feature vectors with semantic and contextual information in review text.Secondly,the attention mechanism is used to assign different weights to review to initialize the feature representation of users and items,and the graph convolutional network is used to capture higher-order collaborative information from the interactive graph of users and items to improve the feature representation,thus predicting the user's rating of items.Finally,the experiment results on the recommendation datasets verify the superior performance of the proposed model.