Research On Recommender Algorithms Based On Latent Factor Model

With the rapid development of Internet,the era of big data has come.However,people can not choose valuable information from the public data,which leads to the problem of "information overload".However the emergence of the recommendation system can very good solve this problem,it can be based on user behavior make reasonable recommendations for the user,the history of recommendation algorithm is the soul of the recommendation system,the latent factor model is the most common recommendation algorithm model of the matrix factorization,since competition born Netflix recommendation algorithm,this algorithm have been scholars love.However,the sparse data greatly reduces the performance of the traditional cryptic meaning model.In addition,the traditional cryptic meaning model needs to traverse the entire training set during the training,resulting in a long optimization time for the model.In order to solve the problems caused by the recommendation algorithm of traditional latent factor model,this paper optimizes the model from two dimensions.Firstly,by optimizing the algorithm itself,the influence of data sparsity on the accuracy of the algorithm is effectively solved.Secondly,the training mode of the algorithm is improved,which greatly reduces the training time of the model.The main research contents of this paper are as follows:(1)In order to solve the data sparse on accuracy of the impact of this algorithm,this paper from the positive and negative samples we do to improve on this factor,through clustering technology to users with similar interests together and for each user group of the popular project to sort,select the project itself had a behavior as negative samples for training.Finally,experiments show that the improved algorithm has excellent accuracy in a sparse data set.(2)Through the training mode of traditional latent factor model were analyzed,and found that the each iteration the need to traverse the entire data set,this optimization time and impact model and the accuracy of the algorithm in a large-scale data set will be greatly decreased,in order to solve the above problem,this paper puts forward the incremental sample way of training,only need to traverse the part of the data set,each iteration optimization can not only shorten the time,but also can recommend to keep good performance in large-scale data set.But the training way still exist insufficient,in this paper,the advantages of integration of above two kinds of training methods,adding momentum factor of its inherited the previous iterative gradient values,in different size Movie Lens data set on the experiment data,the experimental results verify the effectiveness of the improved training method can not only reduce the optimization time of algorithm,but also can improve the accuracy of the algorithm.(3)By integrating the improved algorithm and the improved training method,a hybrid model is proposed,and the experiment proves that the hybrid model has better performance than the single model.