密度敏感模糊核最大熵聚类算法

Density-sensitive fuzzy kernel maximum entropy clustering algorithm

中文关键词  聚类  相对密度  最大熵聚类算法  鲁棒性

英文关键词  clustering  relative density  maximum entropy clustering algorithm  robustness

中文摘要

      提出一种密度敏感模糊核最大熵聚类算法. 该算法首先通过核函数将原始非线性非高斯的数据集转化为 核空间数据集, 然后利用核函数的相似性抵消不属于该聚类的样本数据在聚类过程中对聚类中心求解的干扰, 消除 正则化系数对聚类结果的影响, 进而抑制传统最大熵聚类算法的趋同性. 最后通过引入相对密度项, 解决因样本数 据在特征空间的分布差异而导致的聚类中心求解偏差问题, 从而提高聚类结果的准确性. 实验部分, 本文讨论了算 法参数间的关系以及对聚类结果的影响. 通过与传统模糊C均值聚类算法、核模糊C均值聚类算法、最大熵聚类算 法、最大熵规范化权重核模糊C均值聚类算法以及其他两种改进最大熵聚类算法的聚类结果进行对比分析, 结果表 明本文提出的密度敏感模糊核最大熵聚类算法的聚类性能明显优于其他算法.

英文摘要

      In order to solve the clustering problem of nonlinear non-Gaussian datasets, a density-sensitive fuzzy kernel maximum entropy clustering algorithm is proposed. The algorithm firstly transforms the nonlinear non-Gaussian dataset into kernel-space dataset through a kernel function, and then uses the similarity of the kernel function to cancel the interference of sample data which do not belong to the clustering on the solution of the clustering center in the clustering process. This may be helpful to eliminate the influence of the regularization coefficient on the clustering result, and further inhibit the convergence of the traditional clustering algorithm. Finally, the relative density term is introduced to solve the deviation problem of clustering center solution caused by the difference of sample data distribution in feature space, thus improving the accuracy of clustering results. In the experimental part, the relationship between the algorithm parameters and the influence on the clustering results are discussed. By comparing the clustering results with the fuzzy C-means clustering algorithm, the kernel fuzzy C-means clustering algorithm, the maximum entropy clustering algorithm, the maximum entropy normalized weight kernel fuzzy C-means clustering algorithm, and other two modified maximum entropy clustering algorithms, the clustering performance of the density sensitive fuzzy maximum entropy clustering algorithm proposed in this paper is obviously better than other algorithms.