Research On Collaborative Filtering Algorithm Based On Neural Network And Double Clustering

With the rapid development of Internet technology,the amount of network information increases geometrically,which brings great convenience to daily life of people.However,at the same time,it also brings the problem of information overload to people.In this case,recommendation systems came into being,personalized recommendation systems get the user interested content from massive Internet information and recommend the content to users according to their interest preferences,needs and other information.Recommendation system plays an important role in alleviating the problem of information overload.It has become an indispensable technology application in the Internet era and widely used in various fields.As the core of recommendation system,recommendation algorithms have attracted the attention and research of academia and industry in recent years.Collaborative filtering algorithm is the most widely used recommendation algorithm.With the gradual expansion of users and items,the sparsity of rating data further increases.Neighborhood-based collaborative filtering recommendation algorithms measure the similarity between users or items by the similarity of the rating vector.The algorithms have strong interpretability and is easy to implement.However,in the case of the sparsity of rating data,the performance of rating prediction is not good,and there is still a large space for improvement.As a model-based collaborative filtering algorithm,Matrix factorization has good scalability.Because of its good recommendation effect,it has been widely concerned.The traditional matrix factorization technology complements the original ratings by the inner product of users' latent features and items' latent features,and then makes rating prediction.However,such a simple linear interaction function as matrix inner product cannot capture the deeper and more complex latent features of users and items.In addition,unlike the Neighborhood-based collaborative filtering algorithm,the matrix factorization technology does not have good interpretability and cannot associate the feature dimension with the concept in real life,so it can only be understood as the latent semantic space.With the continuous development and breakthrough of deep learning technology in many fields such as computer vision,speech recognition and natural language processing,more and more researchers pay attention to applying deep learning technology in recommendation field,which not only brings new opportunities but also challenges to researches of recommendation algorithm.In view of the above problems,this paper constructs a rating prediction model based on matrix factorization combined with deep neural network,and improves the traditional collaborative filtering algorithm based on this model.The main work is as follows:First,a matrix factorization model based on deep neural network is proposed.The recommendation model based on matrix factorization is limited by the simple use of linear interaction function-matrix inner product to model the interaction between users and items.In view of this disadvantage,on the basis of the idea of the matrix factorization algorithm and combined with the powerful nonlinear learning ability of neural network,the deeper,more complex and more abstract latent feature representations are obtained from the interactive information of users and items.Also,the model combines the different feature information obtained by each layer of neural network for rating prediction,to a certain extent,it avoids the impact of information loss caused by the feature transformation of neural network.Experiments on several open source datasets show that the model can accurately obtain the latent feature of users and items,and improve the accuracy of scoring prediction.Secondly,a collaborative filtering algorithm based on double clustering of users and items is proposed.The accuracy of traditional collaborative filtering algorithm is low when the rating data is sparse.To solve this problem,DeepMF model is employed to fill the original rating matrix.In order to avoid the premature convergence of traditional K-means clustering algorithm,simulated annealing algorithm is employed to improve K-Means algorithm.The method of similarity calculation is improved by the integration of rating similarity and feature similarity to avoid the problem that a single rating similarity cannot describe the similarity relationship of users or items accurately,and further improve the accuracy of similarity calculation.Then,the model clusters from two directions including users and items,and combines the completion rating matrix to calculate the prediction rating.Experimental results show that the proposed recommendation algorithm effectively reduces the impact of data sparsity,and further improves the accuracy of rating prediction compared with the traditional recommendation algorithm and DeepMF model.