基于功效散度和成对约束的半监督聚类算法  

Semi-supervised clustering based on power-divergence and pairwise constraints

作　　者：向思源 金应华[1] 徐圣兵[1] XIANG Si-yuan;JIN Ying-hua;XU Sheng-bing(School of Applied Mathematics,Guangdong University of Technology,Guangzhou 510520,China)

机构地区：[1]广东工业大学应用数学学院

出　　处：《佛山科学技术学院学报（自然科学版）》2019年第4期48-59,共12页Journal of Foshan University(Natural Science Edition)

摘　　要：现有成对约束半监督聚类算法(CE-sSC)克服了极大熵聚类(MEC)算法不能利用样本成对约束信息的缺点,但CE-sSC算法的惩罚项中各熵项之间相互干扰,不利于惩罚项系数的选择。为克服此问题,基于相对熵提出了一类新的半监督聚类算法(PD-sSC),并把表示成对约束样本信息(外部信息)的相对熵项推广到了功效散度(PD)族。此时,PD指标可取任意的实数,当成对约束数较少时,可通过调整PD散度指标来选择比对比算法表现更好的PD-sSC算法。实验结果显示了PD-sSC算法的优良性质,PD-sSC算法惩罚系数的选择也比CE-sSC算法简单且高效。The classical maximum entropy clustering (MEC) cannot use pairwise constraints information.The existing semi-supervised clustering (CE-sSC) could overcome this weak-point of MEC.However,the entropy terms of penalty term in CE-sSC are twisted together and this would increase the difficult of choice for penalty term coefficients.In order to overcome this issue,this paper proposes a new semi-supervised clustering (PD-sSC) method based on relative entropy and extends relative entropy which depicts the information contained in pairwise constraints to power-divergence (PD) family.PD index could be any real number.And when the number of pairwise constraints is small,one can adjust the PD index with aim to choose some PD-sSC whose clustering performance is better than compared algorithms.Experiment outcome shows that PD-sSC has good clustering performance and its choice of penalty term coefficients is much simpler and more effective than CE-sSC.

关 键 词：半监督聚类 功效散度 成对约束 极大熵聚类 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]