Research On Personalized Item Recommendation Based On Deep Learning

With the rapid development of network technology and e-commerce today,more and more online services have appeared on the Internet,which bring a lot of convenience to people.At the same time,it also leads to a sharp increase in the amount of information on the Internet,and users have to spend more time to choose what they are interested in among a large amount of information,which is the problem of information overload.Recommendation system is an intelligent decision making system that adopts machine learning technology to recommend relevant contents to users based on the analysis of users' interests and preferences.It can help users find appropriate contents within a reasonable time and has been proved to be an effective means to deal with the problem of information overload.At present,although the recommendation system has been widely used in various social fields,there are still many problems in the existing recommendation system.For example,data sparsity,cold start,long tail items are difficult to be recommended,and graph structure data cannot be effectively processed,resulting in low recommendation performance and poor user experience,which restricts the development of personalized recommendation system.Through a deep,nonlinear,automatic feature learning network structure,deep learning can obtain the deep feature representation of users and objects,which has the advantages that traditional machine learning cannot compare.In recent years,with the rapid development of the theory and technology of deep learning,the study on personalized recommendation on deep neural networks has been paid more and more attention by the industry and academia.How to use the principles and techniques of deep learning to alleviate and overcome the problems in the existing personalized recommendation system,so as to improve the performance of the recommendation system,is a topic worthy of research.Based on the principles and techniques of deep learning,this paper studies the data sparsity,the cold start of items,the difficulty in recommending long-tailed items and the inability to effectively process the graph structure data in the existing recommendation system.The main research work and results are as follows:(1)Aiming at the problem of data sparsity in the existing personalized recommendation technology,by using Convolutional Neural Networks(CNN),which has the advantage of strong ability to capture Local features,a Local Similarity Prediction model—LSPCNN is proposed to improve the Prediction of Local Similarity of CNN by adding an adjustment layer.Based on the iterative adjustment of the initial user-item rating matrix,the new model makes users' interests and preferences locally characteristic,and then integrates CNN to predict the missing score,so as to implement personalized recommendation.The experimental results show that the MAE values of LSPCNN model under different data sparsity are lower than those of the existing recommendation methods,which alleviates the sparsity of data in the recommendation process to some extent.(2)In the research on cold start recommendation of existing items,users' temporal preference is rarely considered,resulting in low accuracy of recommendation and poor user experience.A personalized recommendation model of items cold start—Text Rank-LSTM based on Text Rank and long and short term memory network(LSTM)is proposed.In this model,the improved Text Rank algorithm of average information entropy is firstly introduced,and the item feature keywords are extracted from the content information of new item or the comment information of user behavior item.Then,the influence factors of users' positive and negative behaviors are introduced to improve LSTM gate logic,and the characteristic keywords of related items after vectorization are taken as the input of improving LSTM according to the time series of users' behaviors,so as to obtain the knowledge sequence of target users' preference experience.Finally,textrank-lstm personalized recommend-ation model is constructed to implement the recommendation of items.The experimental results show that the textrank-lstm model can alleviate the cold start of new items to some extent while improving the recommendation accuracy,recall rate and novelty.(3)To solve the problem of item long tail in Existing personalized recommendation technology,In particular,due to the small number of evaluations on unpopular items,the existing recommendation algorithm makes it difficult for users to pay attention to long-tail items,which affects the sales of long-tail items.In this paper,an improved deep restricted Boltzmann long tail-distribution recommendation method based on multiple themes—LDBM(Long Tail Items Recommendation Model Based on DBM)is proposed.This method builds the relation between the user and the item by mining the topics feature of item information and the user's preference topics feature,and predicts the unknown and potential preference topics of the user by improving the restricted boltzmann machine,so as to improve the recommendation performance of the long-tail item.The experimental results show that the LDBM model has a certain ability to recommend long-tailed items and improves the effect of personalized recommendation.(4)In view of the existing personalized recommendation technology,which mainly deals with conventional euclidic data such as text and image,it is difficult to deal with the problem of graph structure data.On the basis of the research on Graph Neural Networks(GNN),this paper proposes a personalized recommendation model—GNNLPR(GNN Link Prediction Recommendation)based on the Link Prediction of Graph Neural Networks(GNN).Due to the irregularity of graph data,the disorder variability of graph nodes,and the unequal number of neighbor nodes of each node,personalized recommendation research based on deep learning technology has brought certain challenges.As a kind of run directly on the graph structure of neural network,GNN is the latest development of deep learning theory and technology in graph field.It can capture the dependencies on the graph through the message passing between the nodes of the graph,which has better advantages in analyzing non-euclidean data and recommending different entities and their relationships with each other in the system.GNNLPR model reconstructs the node information by clustering and updating the graph neural network.On this basis,the node information and path information are considered to realize the link prediction recommendation between node pairs in GNN graph.The experimental results show that the GNNLPR model has better results in processing the link prediction and recommendation of the graph structure data,and further improves the performance of the recommendation system.The innovations of this paper mainly include:(1)From the perspective that convolutional neural network(CNN)has a strong ability to capture local features,the data sparsity problem in personalized item recommendation based on deep learning has been studied.By adding an adjustment layer to CNN,a personalized recommendation model LSPCNN has been proposed to improve the prediction of local similarity of CNN.This model changes the deficiency of existing personalized recommendation methods in computing user preference characteristics from global and sparse data,improves the effect of personalized item recommendation,and is innovative to some extent.(2)From the perspective of reflecting the temporal behavior of users' preferences,the problem of item cold-start recommendation based on deep learning has been studied.In the existing studies on cold start recommendation of personalized items,few users' interests or preferences are considered to have time series,resulting in low accuracy of recommendation.This paper has proposed a cold start recommendation model of items--textrank-lstm,which reflects users' time series preference behavior.This model effectively alleviates the problem of cold start of items in personalized recommendation and has certain innovation.(3)From the perspective of GNN,the personalized recommendation problem of link prediction based on deep learning has been studied.In view of the shortcomings of existing personalized recommendation methods,which are difficult to directly process the graph structure data,this paper has proposed a personalized recommendation model based on GNN link prediction—GNNLPR.This model uses the advantages of GNN to aggregate and update operations directly on the graph structure,and fully integrates GNN node information and topology structure information to achieve link prediction and recommendation.