Research On The Improvement Of Collaborative Filtering Recommendation Algorithm Based On Diffusion Filtering On Graph

The recommendation technology is a kind of intelligent technology that obtain the users' interactive data to the projects,analyze the users' interests and recommend the valuable information and data to them actively.One major difficulty of Recommendation system is the sparseness of scoring matrix,and it leads to the collaborative filtering algorithm hard to obtain satisfactory prediction accuracy.Therefore,how to solve the data sparseness of scoring matrix becomes a hot issue of recommendation technology.Given that there is a certain affinity between users(projects),the scoring matrix can be seen as signals on a graph.Therefore,the diffusion filtering on graph can be used to fill the missing items of the scoring matrix.However,how to construct the appropriate graph structure? The existing papers usually takes the cosine similarity of the users(projects)or Pearson correlation as the connection weight between vertices.However,this still does not get rid of the sparsity of the scoring matrix.In order to solve this problem,this paper propose a method which learns the adjacency matrix automatically based on the scoring matrix,its basic idea is to make the learnt adjacency matrix guarantee the most accurate prediction to known scoring matrix.Based on this idea,the learning problem of adjacency matrix is finally transformed to an optimization problem,which can be solved by traditional gradient descent method.Experimental results on two open databases show that the prediction accuracy of this method is significantly higher than that of the existing methods.What we have done in this paper is as follows:1.Because of the problems existing in the DF&CF algorithm,we propose the Gradient Descent improved algorithms based on this algorithm.The experimental results tell us that if we compare the DF&CF algorithm with the improved algorithms,the latter ones can enhance the quality of recommendation obviously.2.Based on the Item-based diffusion filtering algorithm,we propose the User-based diffusion filtering algorithm and combine it with CF and GD.The experiments tell us that compared with original algorithm;the UB-DF algorithm can enhance the effect of recommendation obviously.3.In the experiment,we fill the matrix with Matrix Completion algorithm and run the experiment with the same algorithm.The experiments tell us that compared with the matrix without filling,the filled matrix express better recommendation effect.