Research On Hybrid Recommendation Model Based On Deep Learning

Under the rapid development of the Internet,information overload?unclear user needs are becoming more and more serious.The recommendation system is an important system tool in the Big Data environment to make information distribution accurate,timely,and efficient.Collaborative filtering recommendation is widely used in various scenarios due to its simplicity and efficiency.However,as the number of users and items increases,the interaction data between users and items becomes more and more sparse,affected by the cold start,the performance of the recommended system gradually decreases.With the development of deep neural network technology,deep learning has demonstrated excellent ability to represent heterogeneous data.Applying deep learning to recommendation algorithms has become one of the hot topics of research today.Aiming at the sparseness of the scoring data and the cold start phenomenon in the collaborative filtering algorithm,this paper analyzes the research status of deep learning applied to recommendation algorithms,discusses how to better use the advantages of deep learning to make up for the shortcomings in recommendation algorithms.Two hybrid recommendation models based on deep learning are proposed from different perspectives.The main research work is as follows:1.This paper proposes a hybrid recommendation model based on deep learning that called HRS-DC.Starting from the perspective of auxiliary information,using deep neural networks to extract deep features in user and item attribute information;using convolutional neural network with attention mechanism to mine text features in item text information.The item feature and user feature are input into a neural collaborative filtering model to fit the nonlinear relationship between the user and the item for predicting the score.2.A collaborative filtering recommendation model called cycle-CFAAE based on improved adversarial autoencoder is proposed.Starting from the perspective of deep collaborative filtering,the autoencoder is optimized,a "encoding-decoding-encoding" structure is proposed to reconstruct the potential characteristics of users and items,enhancing the antiinterference performance of the model;the discriminative network and the generated network conduct game training on the scoring vector to improve the representation ability of potential vectors for users and items;and then perform Top-N recommendation through neural collaborative filtering.3.The above two models are verified in the MovieLens,AIV,and Netflix data sets,compared with the classic and latest models from multiple aspects,which improves the accuracy of score prediction,can effectively alleviate data sparsity and cold start,and enhances the top-N recommendation performance.