Research On Recommendation System Based On Deep Learning

The recommendation system,as a method to solve the "information overload" problem brought about by the rapid development of the Internet,saves users the time and effort spent searching for information.The collaborative filtering method is currently one of the most widely used recommendation methods.This topic uses the probability matrix decomposition method in the collaborative filtering method,because the error between the predicted score and the actual score is smaller,and the recommendation is more accurate,but there is also a cold start And the problem of data sparseness.With the in-depth research of deep learning,its advantages of deep mining of item characteristics are gradually used in recommendation systems to alleviate these two problems.The main research contents of this topic are as follows:In order to improve the shortcomings of traditional recommendation methods and retain their advantages in predictive scoring,this topic uses a hybrid method of stacked denoising autoencoders and probabilistic matrix factorization for recommendation.According to the probability matrix decomposition in the traditional collaborative filtering recommendation method,the score matrix is processed to obtain a user matrix and an item matrix.Then use the stacked noise-reducing autoencoder in the deep learning method to train the item matrix to dig deep-level item features to lay the foundation for predicting more accurate scores.Then multiply the user matrix composed of the user latent vector with the item matrix composed of the item latent vector processed by the stacked noise reduction autoencoder,and the inner product of the vector is the predicted score,and the predicted score will rank the top 10 products Recommend to users.Finally,compare the improved hybrid recommendation method of this subject with other recommendation methods,and analyze the advantages.The experiment uses open source data sets for verification,and accuracy and root mean square error are used as evaluation criteria.The experimental results show that the hybrid recommendation method used in this topic has higher accuracy and smaller error than the traditional recommendation method or the noise reduction autoencoder method.It proves that the hybrid recommendation method combined with the probability matrix decomposition method can alleviate the cold start.At the same time,it can accurately predict the characteristics of the score and the stacked noise reduction autoencoder to dig deeper features of items and optimize the advantages of hidden vectors,which improves the recommendation effect.