Research Of Recommendation Algorithm Based On Restricted Boltzmann Machine

With the rapid development of Internet and information technology, the Internet provides more and more information and services for users. However, in the face of the exponential growth of information resources in the Internet, the user is difficult to locate the useful information for them, which will cost a lot of time, this is the problem of information overload. Recommender system is aim to solve this problem, Recommender system can provide personalized recommendation service based on the preferences of different users.Currently, collaborative filtering has become the most used approaches for recommender system. However, collaborative filtering algorithm and recommender system face the challenges of data sparsity、scalability、cold start and so on. Data sparsity is one of the most crucial challenges for the collaborative filtering algorithm or recommender system. Data sparsity will lead to the low prediction accuracy of collaborative filtering algorithm, which will seriously reduce the user experience. Therefore, to solve the problem of data sparsity has great significance to improve the accuracy of collaborative filtering algorithm. In recent years, with the popularity of social network, social network relationship is becoming more and more important in people’s daily life. Friends’ opinion or view in social network tend to influence our decision, therefore, the use of social relationship in social network can solve the problem of data sparsity.Nowadays, deep learning has made significant breakthrough in many fields. Restricted Boltzmann Machine model occupies a central position in the field of deep learning, which can be used to solve the recommendation problem. In current research, Restricted Boltzmann Machine model for collaborative filtering has some defects:Firstly, rating data need to be converted to a K dimensional 0-1 vector unit, this will make too many parameters in model and more complex training process. Moreover, this kind of transformation method is only valid for integer data, double data can’t be converted; Secondly, Model only uses user rating data, however, there are serious data sparseness problem in user rating data. Data sparsity will affect the recommendation quality of model in a certain extent. Thirdly, we are in an era of big data. In the environment of big data, Restricted Boltzmann Machine model has a large number of parameters, which is an enormous challenge for training model. This thesis mainly focuses on the following contents: 1. On the basis of the existing Restricted Boltzmann machine model for collaborative filtering, a Real-Valued Conditional Restricted Boltzmann Machine (R_CRBM) model is proposed in this paper. In the R_CRBM model, rating data does not need to be converted to a K dimensional 0-1 vector unit. Meanwhile, the training process of R_CRBM model also uses rated/unrated information. The experimental results on Baidu and Epinions datasets show that rated/unrated information helps to alleviate data sparsity problem.2. Combine R CRBM model with social network information, we proposed a method RCRBM NTFMT which combine Real-Valued Conditional Restricted Boltzmann Machine model with Nearest Trusted Friends Based on MoleTrust. The experimental results on Baidu and Epinions datasets show that R_CRBM_NTFMT algorithm is good solutions to solve the problem of data sparsity. Moreover, the experimental results show that the train data more sparse the prediction accuracy of R_CRBM_NTFMT algorithm is better than RBM and S_RBM model.3. A common platform is very difficult to train R CRBM model in big data, therefore, a parallelization scheme based on Spark is proposed in this thesis. The experimental results show that the parallelization scheme has a good scalability, which can solve the problem of big data.