Investigation On Personalized Recommendation Method Based On Deep Learning

Recently,with the rapid development of the Internet and computer technology,numerous data has emerged on the Internet,which provides users with a lot of conveniences.However,due to massive data,it is difficult for users to quickly obtain information that meets their needs and interests,thereby resulting in a serious problem of information overload.To solve this problem,many methods have been proposed,among which recommender system is one of the most effective ones.As a personalized information recommendation service,it can effectively communicate between users and information.It mainly obtains the interests and needs of users by analyzing user data,so as to recommend information that they are interested in to users.However,in the process of the recommendation,due to data sparsity,cold start,the design flaws of the algorithm（e.g.,latent feature initialization）,and the difficulty of processing graph-structured data,etc.,the recommendation accuracy rate is low and the user experience is extremely poor,thus limiting the development of recommender systems.Fortunately,with the recent achievements of deep learning in various applications,such as natural language processing,machine translation and image processing,deep learning models have been used in recommender systems,thereby solving various challenges of traditional recommender systems to improve their performance.In addition,leveraging deep learning techniques in recommender systems has also become more popular.Compared with traditional recommendation architectures,deep learning-based recommendation ones provide better learning capabilities for user permission interaction representations.Therefore,the development of deep learning-based personalized recommendation systems has become a trend.At the same time,how to use the relevant principles and technologies of deep learning to overcome and alleviate the problems existing in the existing personalized recommendation system for improving their recommendation performance,is a topic worthy of research.In view of the above-mentioned problems,such as data sparsity,random given user and item latent feature vector initial values（i.e.,random or zero initialization）,cold start and difficulty in processing graph structure data,the relevant research has been carried out,and the following solutions have been proposed as follows:（1）Aiming at the data sparsity problem in existing personalized recommendation systems,a cross-domain recommendation method based on transfer learning（TCD-CF）is proposed.It mainly uses the characteristics of transfer learning that can apply knowledge or patterns learned in one domain or task to other different but related domains or problems.In other words,its purpose is to transfer the knowledge learned from the movie recommendation model to the book recommendation through the expansion of the codebook,then predict the score of the defects in the book data,and finally alleviate the data sparsity of book data.The experimental results show that the model achieves the purpose of reducing data sparsity to a certain extent,and improves the performance of the recommender system.（2）Aiming at the problem of randomly given initial values of user and item latent feature vectors（such as random or zero initialization）in existing personalized recommendation systems,a deep learning-based initialization recommendation algorithm（DLIR）is designed for trust-aware recommendation.In practice,from an optimization point of view,matrix factorization-based methods are sensitive to the initialization of user and item latent feature matrices,because the minimization process in matrix factorization is non-convex.Therefore,a good initialization can lead to better local minima and improve the efficiency and accuracy of the learning process.Experimental results display the proposed DLIR can learn a good initial value of user and item latent feature vectors,thereby further improving the recommendation accuracy.（3）Aiming at the cold-start problem in existing personalized recommendation systems,a social recommendation algorithm based on knowledge distillation（EAF-SR）is proposed.This method first uses stacked denoising autoencoders to reduce the noise of user feedback data,and generates soft targets from user feedback and user trust information.Then,a knowledge distillation method is used to learn reliable information from the generated soft targets.Finally,the pretrained network and the retrained network are combined to provide the final recommendation for each user.Experimental results indicate that the proposed EAF-SR provides a new solution for learning robust representations from discrete input data common in daily life,and can well solve the user cold-start problem.（4）Aiming at the problem that the existing personalized recommendation methods mainly deal with Euclidean data,but it is difficult to deal with non-Euclidean data of graph structure and cannot well capture the dependencies between the internal data of the recommendation system,a predictive recommendation method based on graph convolution networks（SR-GCN）is proposed.As a neural network that operates directly on the graph structure,the graph convolutional neural network is a new extension form of deep learning theory and technology in the field of graphs.It can model non-European spatial data and better capture the internal dependencies’ relation to the data.In the SR-GCN method,based on the（user-item）and（user-user）bipartite graphs,the graph convolutional neural network aggregation and update method is used.At the same time,the topological relationship between the node information and the node structure in the GCN is considered for making Top-N recommendations,thereby further improving the performance of the recommendation system.Experiments conducted on three datasets（i.e.,Gowalla,Ciao and Amazon）show that the proposed SR-GCN demonstrates good results in processing graph-structured data,and further improves the performance of recommender systems.