基于能量的结构最小二乘孪生参数间隔支持向量聚类  

Energy-based Structural Least Squares Twin Parametric-Margin Support Vector Clustering

作　　者：王顺霞 黄成泉 蔡江海 罗森艳 杨贵燕 WANG Shunxia;HUANG Chengquan;CAI Jianghai;LUO Senyan;YANG Guiyan(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang 550025,China;Engineering Training Center,Guizhou Minzu University,Guiyang 550025,China)

机构地区：[1]贵州民族大学数据科学与信息工程学院,贵州贵阳550025 [2]贵州民族大学工程技术人才实践训练中心,贵州贵阳550025

出　　处：《西北民族大学学报（自然科学版）》2024年第4期12-24,共13页Journal of Northwest Minzu University(Natural Science)

基　　金：国家自然科学基金项目(62062024);贵州省模式识别与智能系统重点实验室2022年度开放课题(GZMUKL[2022]KF03);贵州省省级科技计划项目(黔科合基础-ZK[2021]一般342);贵州省教育厅自然科学研究项目(黔教技[2022]015)。

摘　　要：针对小样本学习中传统的基于平面的聚类算法对噪声敏感、计算耗时、类内信息未得到充分利用等缺点导致算法性能表现不佳的现象,提出了一种新的基于能量的结构最小二乘孪生参数间隔支持向量聚类方法.该算法通过将类内协方差矩阵引入目标函数来获取数据的结构信息,同时将能量因子引入每个聚类参数间隔中心超平面,降低噪声和异常值对算法的影响,并用凹凸迭代过程求解目标函数的优化问题,并在多个合成数据集和真实数据集上进行实验.通过统计测试验证该算法的显著性,实验结果证明了所提算法具有良好的性能.In few-shot learning,the traditional plane clustering algorithm is sensitive to noise,time-consuming in computation,and the information in the class is not fully utilized,which leads to poor performance of the algorithm.A new energy structured least squares twin parameter-margin support vector clustering for few-shot learning was proposed in this thesis.The algorithm introduced within-class covariance matrix into the objective function to obtain the structure information of data.Furthermore,the energy factor was introduced into the center hyperplane of each cluster parameter-margin center hyperplane,so that the influence of noise and outliers could be reduced in the algorithm,and the optimization problem of the objective function could be solved by using the concave-convex iterative process.Finally,experiments were carried out on multiple synthetic datasets and real datasets,respectively.The significance of the proposed algorithm was verified by statistical tests.The experimental results demonstrated that the proposed algorithm had excellent performance.

关 键 词：双参数 支持向量聚类 最小二乘 小样本学习 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]