The Research And Improvement Of Collaborative Filtering Algorithm Based On Deep Learning

The core of the recommendation system is the recommendation algorithm.The collaborative filtering algorithm is favored among all recommended algorithms because of its simplicity,efficiency and stability.It is also the most widely used and recommended technique in the recommended algorithms.However,in the actual application scenario,as users and commodities continue to grow,the scoring matrix in the collaborative filtering algorithm gradually exposes the data sparsity and cold start problems,and also ignores a large amount of information in the comment text.In addition,the collaborative filtering recommendation algorithm only uses shallow features to train the model,which restricts the recommendation performance.In the past five years,with the rise of deep learning,the upsurge of artificial intelligence has been set off.In the field of image recognition,text mining,and speech recognition,breakthroughs have been made,which also brings about the possibility of improving the performance of recommended algorithms.The main work of this paper consists of the following two parts:(1)Aiming at the problems of cold start and data sparsity in the traditional collaborative filtering recommendation algorithm,and the degree of dimensionality reduction in matrix decomposition mainly depends on the shortcomings of prior knowledge lacking flexibility,a collaborative filtering algorithm that incorporates singular value energy is proposed(SVE-CF).Firstly,during the data preprocessing process,the noise rate of the user-item scoring matrix is calculated,the degree of replacement of the scoring matrix is determined according to the noise rate,and the approximate score matrix is obtained by using the singular value energy.Finally,the user in the approximate matrix is decomposed by the matrix factorization model.The goods are mapped into the same space,the interaction between the two is determined,and a ratingbased recommendation is implemented.The SVE-CF model and multiple traditional recommendation algorithms were validated on the common dataset Movie Lens.The results showed that the MSE and MAE measures of the SVE-CF model were reduced by an average of about 3%.At the same time,the impact of data sparsity and the cold start band were effectively mitigated.(2)Firstly,the traditional collaborative filtering recommendation algorithm and the classical deep learning model are studied.Based on this,a new recommendation algorithm model is designed and implemented,and a two-channel CNN recommendation algorithm(CDCNN,Combine-is proposed to integrate user and commodity reviews.Double CNN).First,the user and product review texts are vectorized into word vectors,and then the user and the items are extracted by using two CNN networks respectively,and then the abstract features of the user and the items are mapped to the same feature space through the dot product in the shared layer,and finally Predict the user's rating for a particular item.After experiments on the public datasets of Amazon,Yelp and Beer,the results show that the MSE of the model on different datasets is smaller than other benchmark algorithms,and the data sparsity problem is effectively alleviated.