Research On Rating Prediction Of Recommendation System Based On Deep Learning

In recent years,the recommendation system has been widely used in B2C network sales,social network services,search engines,and positioning services,and has achieved good results,but the recommendation system still has some constraints,such as data sparsity.At present,most recommendation system rating prediction work uses deep learning technology,first based on the existing user-item rating matrix,and then use the auxiliary information of users and items to solve the above problems,although it can make up for the problem of data sparsity to a certain extent,However,there is a problem of ignoring context semantics and syntactic information,which affects the recommendation performance.In addition,most of the popular rating prediction models use matrix decomposition for fusion,but the initial implicit feature matrix of users and items has a great influence on matrix decomposition,and some Work to avoid or ignore this problem.In view of the problems in the above research work,the thesis conducts an in-depth study on the rating prediction algorithm of the recommendation system based on deep learning,and proposes a novel dual autoencoders matrix factorization model(DAE-MF),which fully Use deep learning techniques,scoring matrices,and user project comment information to improve recommendation accuracy.The main contents of this article are as follows:(1)A Convolutional AutoEncoder(CAE)is proposed to process project review information.The convolutional autoencoder models the document review information through unsupervised pre-training,which can not only extract the potential features in the review information,but also fully consider the impact of word sequences and context information on the potential features of the extracted items.(2)Stacked Denoising AutoEncoder(SDAE)is selected to reconstruct the user's best initial value.The network learns hidden layer features through layer-by-layer unsupervised learning,and uses the hidden layer features as the user's initial value,which effectively avoids the effects of random or zero-initialized user potential features on matrix decomposition.(3)A Dual AutoEncoder Matrix Factorization model(DAE-MF)is proposed which integrates the matrix decomposition of a convolutional autoencoder and a stacked noise reduction autoencoder.In the case of extremely sparse data,the DAE-MF model can improve the user's score prediction of the item.(4)A large number of experiments were performed on three real data sets to verify that the model proposed in this paper outperforms existing work in rating prediction tasks.