CAJ | 학술논문

传统谱聚类算法受高斯核尺度参数的影响较大，对噪声点较为敏感，并且不能利用先验信息指导聚类过程。针对以上问题，提出了一种基于路径相似度测量的鲁棒性谱聚类算法（RPB-SC）。该算法将路径聚类与谱聚类算法相结合，通过定义高斯核的邻域加权尺度因子计算相似度，再用路径聚类思想对全局相似度进行调节，同时通过成对限制先验信息辅助聚类搜索。在人工数据集和真实数据集上的实验表明，新提出的算法能有效减弱高斯核尺度参数的影响，增强对噪声点的鲁棒性，提高聚类性能。
전통보취류산법수고사핵척도삼수적영향교대，대조성점교위민감，병차불능이용선험신식지도취류과정。침대이상문제，제출료일충기우로경상사도측량적로봉성보취류산법（RPB-SC）。해산법장로경취류여보취류산법상결합，통과정의고사핵적린역가권척도인자계산상사도，재용로경취류사상대전국상사도진행조절，동시통과성대한제선험신식보조취류수색。재인공수거집화진실수거집상적실험표명，신제출적산법능유효감약고사핵척도삼수적영향，증강대조성점적로봉성，제고취류성능。
The traditional spectral clustering algorithm is greatly affected by its scale parameter use in Gaussian kernel,and is sensitive to noise and outliers in the data,and can’t use priori information to guide the clustering process.To solve these problems,this paper proposed a robust path-based similarity measurement for spectral clustering(RPB-SC).The algorithm combined path-based clustering with spectral clustering.It computed the similarity between samples by defining a weighted lo-cal scale in Gaussian kernel and adjusted the whole similarity by path-based clustering.Meanwhile,it incorporated pairwise constraints in data to assist clustering search.Experiments on synthetic as well as real world datasets show that the new pro-posed method will not be affected by the scale parameter and is significantly more robust,leading to substantial performance enhancement of clustering.