News Recommendation Technology Based On Deep Learning In Dual-structural Network

With the advent of the age of Big Data,information on the Internet is exploding at an exponential rate.It is increasingly difficult for people to quickly find useful information for their limited needs and get rid of spam.Based on the main structure of the Internet,the dual-structural network adds the secondary structure based on the "radiation replication" model,and implements "deep redundancy" on the current Internet main structure with the innovative ideas of physical change and the two element structure,and provides personalized information services to help users select useful information quickly.The traditional user based collaborative filtering method calculates the similarity between the user and the other users first,and then recommends items that the similar users had browsed.The whole calculation process is heavily dependent on the user item score matrix.If the score matrix is too sparse,it will lead to the inaccuracy of the similarity calculation,and it is difficult to apply directly to the personalized recommendation service of the dualstructural network.Therefore,there exists a difficult problem in the recommendation mechanism of dualstructural network,that is,how to avoid the problem of data sparsity and cold start problem,and produce accurate recommendation information quickly.Aiming at the requirements and characteristics of the dual-structural network,this thesis designs a deep learning based collaborative filtering algorithm for dual-structural network(DLCF-DSN)to help users make personalized information recommendations.A convolution neural network is introduced into the field of information classification.A multi-branch convolution neural network(MBCNN)is designed to classify news information.Based on this,a margialized denoising autoencoder collaborative filtering algorithm(MDACFA)is proposed to be deployed on the edge server.In order to further enhance the recommendation effect,the T-WORD2 VEC recommendation algorithm is proposed to mine individual user interests and hobbies deeply at the user terminal.The details are as follows:1)A multi branch convolution neural network(MBCNN)algorithm for news information classification is proposed to solve the problem that the traditional classification algorithm can not classify the news pages accurately,thus affecting the recommendation effect in the dual-structural network.First,Web pages are preprocessed to extract the text features.Then the HTML feature is extracted from the source code of the news page.After that,HTML features and text features are fused and input into MBCNN algorithm to classify them.MBCNN contains multiple MbcModule.Each MbcModule has five branches.Each branch learns different features and enhances the expression ability of the model.Finally,a training algorithm is proposed for MBCNN algorithm to reduce the value of its loss function,the result of MBCNN classification is encapsulated into UCL and sent to the edge server as the input of the recommendation algorithm.2)A deep learning based collaborative filtering algorithm for dual-structural network(DLCFA-DSN)is proposed,which includes the collaborative filtering algorithm MDACFA deployed on the edge server.The result of MBCNN is used as the input of MDACFA collaborative filtering algorithm.The loss function is defined according to the MDACFA algorithm,and algorithm is trained by fitting the score matrix with items and user attributes.After the training is completed,the scoring information can be calculated only by the properties of the item and the attribute of the user,thus avoiding the effect of the sparsity of the score data,and to a certain extent,the cold start problem is solved.In order to further enhance the recommendation effect,T-WORD2 VEC recommendation algorithm is proposed to further explore individual user interests in user terminals.3)On the basis of the dual-structural network prototype system,a dual-structural network news information recommendation system based on the above algorithm is designed and implemented,and the feasibility and performance of the MBCNN and DLCFA-DSN algorithms are verified on the related data sets.The experimental results show that the MBCNN algorithm can classify the news information effectively,compared with the traditional machine learning classification algorithm,the accuracy is higher.The DLCFADSN algorithm can fully combine the characteristics of the dual-structural network,compared to the traditional recommendation algorithm,the algorithm is less affected by the problem of data sparsity,and the accuracy is higher.