摘要： 聚类分析在数据挖掘领域有着广泛的应用，该文提出一个聚类新思路，它不需要任何参数的假设，只基于数据两两之间的相似性。该方法假设数据点之间存在随机游走关系，根据数据相似性构造随机游走过程的转移矩阵，当随机游走过程进入收敛期后，t阶转移矩阵揭示了数据点的分布。用迭代方法寻找最小的KL-divergence来对这些分布聚类。该方法具有严谨的概率理论基础，避免了传统算法需要参数假设、限于局部最优等不足。实验表明，该算法具有较优的聚类效果。

关键词: 聚类, 随机游走, KL散度

Abstract: Clustering analysis is broadly applied in data mining. This paper presents a new idea in clustering based on pair-wise similarities, and assumes no parametric statistical model. Similarities are transformed to a Markov random walk probability matrix. It is assumed the dataset is under a Markov random walk process. When the process is going into convergence, the t-step transform matrix indicates the distribution of the dataset. It uses iterative algorithm to cluster these data with the goal of decreasing KL-divergence. This method has a solid foundation of probability theory, which can avoid some insufficiency of the traditional algorithms. The experiment shows the algorithm can achieve better results than K-means and mixture models.

Key words: clustering, random walk, KL-divergence

中图分类号: 

• TP301.6