A Research On Collaborative Filtering Recommendation Algorithm Based On Improved K-means And TimeSVD++LR

The amount of information generated by the society is more and more huge.Information overload is becoming increasingly prominent.Collaborative filtering is a common method to reduce information overload in recommendation system.However,most current recommendation algorithms based on collaborative filtering have problems of low scalability and data sparsity.To solve these problems,this thesis proposes a collaborative filtering recommendation algorithm based on improved K-means and time SVD++LR.The research contents are summarized as follows:(1)To solve data sparsity problem,this thesis proposes a time SVD++ and linear regression recommendation algorithm based on learning rate function.First,based on SVD++algorithm,the algorithm model introduces time perception factors.There are two main time effect attributes in benchmark prediction factors: user bias and object bias.They change with time.At the same time,the user's preference for goods is affected by time.After obtaining the time SVD++ algorithm model,the optimization function is derived by the random gradient descent method.The learning rate function is introduced to speed up the iterative algorithm execution speed,to obtain the prediction score and fill the sparse matrix.The feature vector is constructed according to the complete matrix.The original data set is trained by linear regression model to get the regression parameters.The regression parameters and eigenvectors are combined to make the secondary score prediction by using linear regression prediction model.Then the final prediction score fills the user scoring matrix.(2)Time SVD++LR algorithm reduces the problem of data sparsity.However it needs to search the nearest neighbor in the whole data space when it is recommended.That has much calculation and less scalability.To solve the problem of low scalability,this thesis proposes an improved K-means and time SVD++LR collaborative filtering recommendation algorithm.The algorithm uses time SVD++LR to predict and fill the high-dimensional sparse matrix,to reduce the bad impact of data sparsity on the clustering results.Then,the selection method of initial clustering center in K-means algorithm is improved.Then,the improved K-means clustering algorithm is used to cluster on the complete data set.After clustering,the target users are divided into clusters with the highest similarity.The scope of searching the nearest neighbor is reduced to the target cluster.The search space is greatly reduced,as well as amount of calculation.Besides,the scalability of the algorithm is improved.Finally,the collaborative filtering recommendation is carried out in the target cluster to get the Top-N recommendation lists.In conclusion,the improved K-means and time SVD++LR recommendation algorithm proposed in this thesis improves and optimizes the data sparsity and scalability of traditional collaborative filtering algorithm.Data sets in Movielens were selected for experimental verification.The results show the proposed recommendation algorithm performance is better.The recommendation effect is improved by about 1.93%,compared with LFM algorithm.