Research And Application Of Web Community Recommendation Algorithm Based On Deep Learning

With the rapid development of Internet technologies and social networks,more and more people are gathered in the Web community because of their interest,which leads to endless stream of Web communities.The proliferation of Web communities has provided users with more choices than ever before,but too many choices have brought the problem of "information overload".Thus,the recommendation system came into being.Nowadays,deep learning technology has made remarkable achievements in Image Identification,Natural Language Processing and other fields.This thesis will focus on the Web community,and study how to use deep learning to improve the performance of recommendation.In order to apply the deep learning technology to the recommendation system,this thesis selects the classical deep learning model as a breakthrough and proposes two new recommendation algorithms: One is the combination of Restricted Boltzmann Machine(RBM)and the traditional Item-based Collaborative Filtering(ICF)algorithm,and the similarity calculation method used in Collaborative Filtering algorithm are improved;The other is the application of Deep Neural Network(DNN)model,which implements cross-domain(CD)recommendation.The Douban datasets are obtained by customized crawlers and the performance of the two new algorithms under this datasets are evaluated.This thesis firstly introduces the research background of the topic,illustrates the current development of Web community recommendation,and points out the significance of applying deep learning to the recommendation system.The concepts,algorithms and evaluation indicators commonly used in recommendation system are elaborated afterwards,and the basic principles of neural network and deep learning are briefly summarized.On the basis of relevant theoretical analysis,a new Web community recommendation algorithm based on RBM-ICF is proposed by using the idea of "candidate generation and ranking" in the YouTube recommendation system,and the overall framework of the algorithm,the implementation process and the improved similarity calculation method are described in detail.Meanwhile,referring to the idea of cross-domain recommendation,a CD-DNN based Web community recommendation algorithm is used to introduce cross-domain information to recommendation system.Then,inspired by the idea of “expert users”,this thesis implements a web crawler based on the Douban Group and proposes the idea of crawling the “Douban Expert” data to alleviate the problem of data sparsity that exists when constructing a user-item matrix in massive data.Finally,the two recommendation algorithms mentioned in the thesis are evaluated under the datasets of Douban Group.The experimental results show that the RBM-ICF algorithm can significantly improve the coverage and novelty of recommendation under the premise of sacrificing accuracy accurately.The CD-DNN algorithm has obvious advantages in solving the problem of cold start,and the algorithm can effectively enhance the accuracy of recommendation.