Rating Prediction Algorithm Based On Collaborative Filtering

With the explosive growth of the data in the internet,such as the data of users,commodity,transaction records,social information,the large scale information resources have floored into the internet,and it can easily cause the phenomenon of"information overload".To solve this problem,personalized recommendation technology came into being,it can provide users with their own characteristics of information services and decision support.Collaborative filtering algorithm is a popular problem in personalized recommendation technology.It analyzes the user behavior,digs other users with similar interests to the specified user in the user group,synthesizes these similar users to rate the target items,and forms the recommendation module.Target item rating prediction.However,with the increasing scale of data,collaborative filtering recommendation algorithm is also facing a series of challenges,such as data sparsity problems,scalability problems,recommended accuracy issues.This thesis studies the sparsity and extensibility of data in collaborative filtering algorithm based on model,and concentrates on the recommendation algorithm based on the restricted Boltzmann machine and matrix singular value decomposition.The main tasks include:Firstly,the research on the architecture and development of the traditional collaborative filtering algorithm is introduced,and the algorithm of collaborative filtering based on neighbor similarity and model based is introduced in detail.Depth study of RBM model network structure,contrastive divergence training method,a detailed analysis of the S VD model of the theoretical method,the implicit latent factor model and regularization method were described in detail.Secondly,the RBM-based collaborative filtering recommendation algorithm is improved to add the behavior information that the user has browsed but not rated in the training data to form the collaborative filtering based on the conditional restricted Boltzmann machine prediction algorithm,and the existing CRBM for users to improve,put forward the items for the CRBM model.The experimental results show that the improved CRBM algorithm is better than the CEBM collaborative filtering algorithm for users.Thirdly,the SVD ++ prediction model based on the user behavior attribute is analyzed and improved.The potential information of the user's historical behavior records is added.This thesis proposes an improved Asymmetric SVD and its dual model,in which the user characteristic vector matrix in the original SVD model is replaced with the hidden feature vector matrix containing the user's preference.And the ASVD prediction model is also extended to join the k-nearest neighbor algorithm to form a hybrid recommendation model for rating prediction.The experimental results show that the proposed hybrid model can effectively improve the prediction accuracy of the recommendation system.