Research And System Implementation Of News Recommendation Technology Based On Graph Neural Network

Personalized recommendation technology has been developing rapidly in recent years.It has been widely used in the field of news recommendation and achieved good results.Personalized news recommendation technology is a customized news recommendation based on user preferences.It calculates the Top-k ranking of news based on user preferences by analyzing a series of historical behaviors and attributes of users,and recommends news that users may be interested in.The key in personalized news recommendation is the expression of user and news features.However,it is difficult to quantify user interest accurately.To this end,this thesis carries out research on user interest modeling,and proposes a news recommendation network model based on users’ long-term and short-term interests.The model is divided into two modules,a graph convolutional network module that characterizes users’ long-term interests and news features,and a neural network module that characterizes users’ short-term interests.The model calculates the user features and news features obtained from the two modules through a fully connected neural network,and finally obtains the probability that the user clicks the candidate news.The main work is as follows:Firstly,the design of the Feature Fusion Graph Convolutional Neural Network.Aiming at the user-news interaction,a graph convolutional neural network for feature fusion is proposed.The module enhances the final output of the graph network by feature fusion method.The most useful feature information is retained by attention calculation,and these features are fused to enhance the user’s long-term features.The fused features are used as the final output to obtain an enhanced node embedding representation,which improves the expressive ability of the model.Then,the design of the User Short-term Interest Module based on Weight Activation Unit.Aiming at the dynamic change of users’ short-term interests,a module of users’ short-term interest based on weighted activation unit is proposed.Firstly,the module uses the input obtained from the basic feature extraction module to calculate the content level user embedding and the sequence level user embedding,then calculates the corresponding weights of the user behavior features through the weight activation unit,and finally calculates the user short-term interest embedding by weighting.Finally,implementation of news recommendation system.Design and implement a personalized news recommendation system,which realizes the basic functions of users’ personalized news recommendation,news reading and so on.By applying the framework of this thesis to the recommendation system,the accuracy of personalized news recommendation is improved.