网络推荐系统中K-均值聚类算法的研究及优化

Research and Optimization of K-Means Clustering Algorithm in Network Recommendation Systems

研究了聚类分析技术在网络推荐系统中的应用。由于k均值(k-means)聚类算法易受到局部最优与噪声点等因素的影响,文章结合DBSCAN(Density-based Spatial Clustering of Application with Noise)算法和MMD(Max-Min Distance)初始聚类中心选取算法,对原始k-means算法进行了改进,提出了DMK(Density-based and Max-min-distance K-means)算法。该算法使用DBSCAN选取高密度点作为第一个聚类中心点的备选范围,接着选择相距最远的K-1个点作为其余的K-1个聚类中心,然后用得到的这组初始聚类中心进行k-means聚类。仿真与实验结果表明,该算法选择的初始聚类中心比较分散且代表性好,聚类的迭代次数减少,聚类结果的纯度提高。

In this paper, we study the application of clustering analysis in network recommendation systems. The k-means clustering algorithm is susceptible to local optimum and noise. This paper improved original k-means algorithm combined with DBSCAN and MMD algorithm and proposed a DMK algorithm. The DMK algorithm used DBSCAN to select high density points as the first cluster center's selection range, then used MMD algorithm to select the rest k-1 points that are farthest away from each other as the rest k-1 cluster centers. These k points are used as initial cluster centers to continue k-means clustering.The results of simulation and experiment showed that the DMK algorithm selected dispersed and representative initial cluster centers, the number of iterations decreased and the clustering results' purity raised.