Research On Personalized Recommendation Algorithm Based On Deep Learning

With the rapid growth of Internet information,we are in an era of information overload.How to alleviate information overload has becoming a development problem.Personalized recommendation technology is one of the effective tools to solve these challenges and bridge the gap between users and information.The algorithm of the original recommendation system provides all users with the same recommendation information,and does not make personalized recommendations based on users' preferences.All users will only get the same information recommendation.With the increase in the amount of information on the Internet,the personalized needs of users are increasing,so the introduction of personalized recommendations is of great significance.Collaborative filtering algorithm is a recommendation algorithm used in an extensively range of research systems.However,due to the increasing number of users and projects,the data sparsity also increases,the collaborative filtering's cold boot problem becomes more obvious,and the algorithm is greatly restricted.Deep learning has a powerful representation ability in learning abstract feature representation.Therefore,on the basis of existing research,this thesis combines deep learning with recommendation algorithms,and uses the characteristics of deep learning to have good feature extraction capabilities to solve the data sparseness and cold boot problems in recommendation algorithms,and explore the deep features of the data set.Improve the accuracy of the recommendation algorithm.Aiming at the phenomenon of data sparsity and cold boot in collaborative filtering algorithm,a new matrix factorization recommendation model is proposed based on matrix factorization algorithm.Use the convolutional neural network in deep learning to extract auxiliary information of users and items,add the auxiliary information to the original feature matrix,and use the deep neural network to extract features of the input information.Finally,the inner product operation is performed according to the feature matrix added with auxiliary information,and the prediction score is obtained.Aiming at the problem that traditional matrix factorization algorithms cannot learn the deep non-linear characteristics between users and items,which leads to the poor accuracy of Top-N recommendation,a composite recommendation model based on dual variational confrontation autoencoders is proposed.The dual autoencoder model is used to simultaneously learn the low-dimensional feature representations of users and items to reduce the time complexity of the method.At the same time,the variational autoencoder and the confrontation generation network are combined to improve the reconstruction target of the variational autoencoder.And added additional attributes about users and items to alleviate the cold start problem.Finally,when the algorithm converges,the user's feature matrix is multiplied by the learned item to obtain the prediction result.In the end,to verify the effectiveness of the two models proposed in this thesis,the thesis compares the proposed model with PMF,CDAE and other models on the Movie Lens and Netflix datasets based on multiple evaluation indicators.The experimental results show that the proposed model is based on The collaborative filtering model improved by deep learning and the adversarial autoencoder model based on deep learning are effective and feasible.