Research On Collaborative Filtering Method Based On SVR

The collaborative filtering methods still dominate the recommendation system,which provide precise message for users,and provide better using and buying experiences for users,so that become crucial beneficial means for commercial business.Collaborative filtering algorithms are needed for users to filter the unnecessary information and make it as much as possible to predict users' interests and behavior pattern according to existing experiences.However,the collaborative filtering algorithms evolves several segmentation area,each of them has their particular advantage and nice application.In consideration of the nice predicting characteristic of support vector regression(SVR)compared with other machine learning algorithms in the traditional research,how could support vector machine applies better in collaborative filtering field is the keynote in this paper.According to the fact that the data sparsity affects the prediction accuracy of collaborative filtering algorithm,a Sparseness Segmentation algorithm is brought forward.First the Support Vector Regression based on iterative prediction is used to estimate the missing scores for the relatively weak-sparse dense data part of sparse U-I matrix,which has the advantage in predicting the high dimensional small sample data.Then the imputative item-based collaborative filtering algorithm is employed to predict the left unpredicted data.The results in several real-world datasets show it is more fitted for the strong sparse datasets' prediction,and the prediction results could be better compared with the item-based collaborative filtering.Except for the method above,to apply support vector machine to larger scale dataset in collaborative filtering,a method called Collaborative Filtering Algorithm Based on Deep Support Vector Autoregression is brought forward according to the fact that neural autoregressive approach and the support vector machine could be well applied in the deep learning field.This method replaces the last layer's active function in neural networks by linear support vector regression function,to build unique autoregressive model for each user.The results in several real-world datasets showthat relying on the fact that Neural Autoregressive Approach's relatively good application on predicting datasets' missing value,the application of support vector regression can further improve the predicting performance by learning a minimum edge based logarithmic loss.After predicting the unrated items for the users,the rest of paper will focusing on managing the predicting results for recommendation results for users using the Top-N recommendation to prove the practicability of proposed methods in this paper.