摘要
为了有效利用少量先验信息提高多视角数据聚类效果,提出一种基于距离度量学习的半监督多视角谱聚类算法(简称ML-SMC)。首先,利用距离度量学习引入先验信息,将多视角数据映射到反映先验约束条件的空间。然后,根据相似性构造每个视角的视图,将多视角聚类问题转化为最小正则割的图划分问题。实验结果表明,MLSMC算法聚类结果的精度优于3种经典的多视角聚类算法和4种半监督单视角聚类算法。并且通过利用少量先验信息ML-SMC算法能够有效提高聚类效果。
In order to take the advantage of prior knowledge to improve clustering performance,based on distance metric learning（ MLSMC）,a semi-supervised multi-view spectral clustering algorithm was proposed. The prior knowledge was incorporated into clustering process by distance metric learning,which mapped data into a new space which subjects to prior knowledge. Each graph of views was constructed according to similarity metric,and then the problem of multi-view clustering was formulated as an optimization problem of minmum normalized cut. Experiments showed that the quality of clustering results of ML-SMC is superior to three classical multiview clustering algorithms and four semi-supervised single-view clustering algorithms,and the precision of ML-SMC could be significantly improved by incorporating some prior knowledge.
出处
《四川大学学报（工程科学版）》
EI
CAS
CSCD
北大核心
2016年第1期146-151,共6页
Journal of Sichuan University (Engineering Science Edition)
基金
国家自然科学基金资助项目(11471001)
关键词
距离度量学习
多视角聚类
谱聚类
半监督聚类
数据挖掘
distance learning
multiview clustering
spectral clustering
semi-supervised clustering
data mining
