结合弱监督信息的凸聚类研究
发布时间：2018-09-19 09:41
【摘要】：基于目标函数的聚类是一类重要的聚类分析技术,其中几乎所有算法均是经非凸目标的优化建立,因而难以保证全局最优并对初始值敏感.近年提出的凸聚类通过优化凸目标函数克服了上述不足,同时获得了相对更稳定的解.当现实中存在辅助信息(典型的如必连和/或不连约束)可资利用时,通过将其结合到相应目标所得优化模型已证明能有效提高聚类性能,然而,现有通过在目标函数中添加约束惩罚项的常用结合方式往往会破坏其原有凸目标的凸性.鉴于此,提出了一种新的结合此类弱监督辅助信息的凸聚类算法.其实现关键是代替在目标函数中添加约束,而是通过对目标函数中距离度量的改造以保持凸性,由此既保持了原凸聚类的优势同时有效提高了聚类性能.
[Abstract]:Clustering based on objective function is an important clustering analysis technique in which almost all of the algorithms are established by the optimization of non-convex objects so it is difficult to ensure the global optimization and be sensitive to the initial value. The convex clustering proposed in recent years overcomes the above shortcomings by optimizing convex objective functions and obtains a more stable solution at the same time. When there are auxiliary information (such as mandatory and / or non-connected constraints) available in reality, it has been proved that the optimization model can effectively improve the clustering performance by combining it with the corresponding target optimization model, however, The commonly used combination of constraint penalty items in the objective function will destroy the convexity of its original convex object. In view of this, a new convex clustering algorithm combining this kind of weakly supervised auxiliary information is proposed. Instead of adding constraints to the objective function, the key is to maintain the convexity by modifying the distance measure in the objective function, which not only preserves the advantages of the original convex clustering, but also improves the clustering performance effectively.
【作者单位】： 南京航空航天大学计算机科学与技术学院;
【基金】：国家自然科学基金项目(61672281)~~
【分类号】：TP311.13
本文编号：2249762
[Abstract]:Clustering based on objective function is an important clustering analysis technique in which almost all of the algorithms are established by the optimization of non-convex objects so it is difficult to ensure the global optimization and be sensitive to the initial value. The convex clustering proposed in recent years overcomes the above shortcomings by optimizing convex objective functions and obtains a more stable solution at the same time. When there are auxiliary information (such as mandatory and / or non-connected constraints) available in reality, it has been proved that the optimization model can effectively improve the clustering performance by combining it with the corresponding target optimization model, however, The commonly used combination of constraint penalty items in the objective function will destroy the convexity of its original convex object. In view of this, a new convex clustering algorithm combining this kind of weakly supervised auxiliary information is proposed. Instead of adding constraints to the objective function, the key is to maintain the convexity by modifying the distance measure in the objective function, which not only preserves the advantages of the original convex clustering, but also improves the clustering performance effectively.
【作者单位】： 南京航空航天大学计算机科学与技术学院;
【基金】：国家自然科学基金项目(61672281)~~
【分类号】：TP311.13
【参考文献】
相关期刊论文 前2条
1 方玲;陈松灿;;结合特征偏好的半监督聚类学习[J];计算机科学与探索;2015年01期
2 刘铭;刘秉权;刘远超;;面向信息检索的快速聚类算法[J];计算机研究与发展;2013年07期
【二级参考文献】
相关期刊论文 前2条
1 倪巍伟,孙志挥,陆介平;k-LDCHD——高维空间k邻域局部密度聚类算法[J];计算机研究与发展;2005年05期
2 吴郢,阎平凡;结构自适应自组织神经网络的研究[J];电子学报;1999年07期
，本文编号：2249762
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2249762.html
