Research On Recommendation Algorithm Based On Improved Clustering And Asymmetric User Similarity

The rapid development of social networks leads to the rapid growth of information data of network, but it has become more difficult for people to efficiently meet the real part of the demand from these massive data, and thus the recommendation system came into being. Recommendation system analyzes and models through the user’s historical behavior data, mining user potential interest, and then recommend goods or information to the user. Wherein the recommendation system algorithm is the most widely used collaborative filtering algorithm, the algorithm predicts the score of the target users to project. There are two flaws on traditional collaborative filtering algorithm which is base on user similarity. One happens when holding that the similarity between users is symmetrical, this assumption can lead to errors in some cases; another occurred in the choice of the user k-nearest neighbor, target users need to calculate and compare with all other users in project ratings, when the number of users is growing, looking for k nearest neighbor computational cost will becomes high. This paper deeply analyzes the existing problems on user similarity measure methods, proposes an improved user similarity measure, further study the clustering algorithm to solve the low efficiency of the k-nearest neighbor selection problem, and proposes an improved K-means clustering algorithm. Finally, we design a new collaborative filtering algorithm using the two improved algorithms. The main work is as follows:1. Elaborates the existing problem of user similarity measure method through instances, proposed an asymmetric user similarity measure, this method can reasonably represent similarity between users. Furthermore, this article describes the common models to solve the data sparsity:implicit semantic model, the model predicts the unknown scores by singular value decomposition. This article combines the asymmetric similarity model and implicit semantic model, designs a collaborative filtering algorithm based on asymmetric similarity. Finally, we have a verification and comparison on the MovieLens and Douban data sets in the experiments, and use the root mean square error and the mean absolute error as the criterion. Experiments show that the method can improve the similarity measure to some extent on the recommendation quality.2. Introduces the traditional K-means clustering algorithm, and illustrates two problems exist in the traditional method through experiments, one is that the number of cluster centers is difficult to determine, frequently rely on the algorithm users familiarity with the field in which the data is located; the other one is due to the random selection of the initial cluster centers, results of multiple clustering is instability, and local optimal solution easily occur. This paper proposes an improved K-means clustering algorithm based on individual contour coefficient, the algorithm can adaptively determine the number of cluster centers, and reasonably distributes cluster centers into the data object. Not only clustering results has improved, but the results of multiple cluster is very stable. In addition, the improved algorithm inherits the advantages of the traditional method of high efficiency and easily to implement. Finally, we use this algorithm for collaborative filtering algorithm which is previously proposed to improvement, proposed collaborative filtering algorithm based on asymmetric similarity and the improved K-means. The improved recommendation algorithm greatly reduces user search range in determining the user k-nearest neighbor, be able to determine the nearest neighbor set quickly, recommendation efficiency has also improved. Experiments compare several other recommendation algorithms, and confirm the proposed recommendation algorithm has a better recommendation quality.