期刊文献+

基于自然邻的自适应谱聚类算法 被引量:3

An Adaptive Spectral Clustering Algorithm Based on Natural Neighbor
下载PDF
导出
摘要 在传统谱聚类算法中,构造相似矩阵时需要人为输入尺度参数;除此之外,之后的k-means过程中还需要人工输入确切的聚类数目,而以上两个参数对聚类效果影响巨大。针对以上问题,提出了一种基于自然邻的自适应谱聚类算法。该算法不需要人为输入任何参数,完全实现自适应,主要方式是通过自然邻算法获取各点之间的邻近信息,其中包括自然邻个数、自然逆邻个数、自然邻居集以及自然逆邻居集。通过实例分析,在多重尺度数据集下或者在流行数据集中,充分利用以上先验信息构造出更加符合实际情况的相似矩阵。另外,根据近邻传播思想获得聚类数目。将该算法运用于部分人工数据集上,且与谱聚类算法进行比较,聚类效果显著改进。实验结果表明,该算法具有一定的有效性和优越性。 In traditional spectral clustering algorithm, the input scaling parameters are needed to construct the similar matrix. In addition, the exact number of clusters is needed to be input in the subsequent k -means process. The above two parameters have a huge influence on the clustering effect. Aiming at the above problems,an adaptive spectral clustering algorithm based on natural neighbors is proposed. It does not need to input any parameters artificially and can achieve complete self-adaptation,the main way of which is to obtain the prox- imity information between the points by the natural neighbor algorithm, including the number of natural neighbors and inverse natural neighbors, the natural neighbor sets and the inverse natural neighbor sets. Through the case analysis, in the multi-scale or popular data set, the above priori information is made full use of to construct a similarity matrix more consistent with the actual situation. In addition, the number of clusters is gained according to the idea of spread of neighbors. The algorithm is applied to some artificial data sets, and compared with the spectral clustering algorithm, improving the clustering effect remarkably. Experimental results show that it has certain validity and superiority.
出处 《计算机技术与发展》 2017年第11期19-23,共5页 Computer Technology and Development
基金 重庆市基础与前沿研究计划项目(cstc2013jcyj A40049)
关键词 谱聚类 自然邻 自适应 尺度参数 聚类数目 spectral clustering natural neighbor adaptive scaling parameter number of clustering
  • 相关文献

参考文献11

二级参考文献122

共引文献349

同被引文献14

引证文献3

二级引证文献12

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部