基于自然邻居局部中心搜索的粒球生成算法  

Granular Ball Generation Algorithm Based on Local Center Search of Natural Neighbors

作　　者：张璐 邵亚斌 李晨 何怡 ZHANG Lu;SHAO Yabin;LI Chen;HE Yi(Department of Mathematics and Physics,Jinzhong College of Information,Jinzhong 030800,China;School of Science,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]晋中信息学院数理教学部,晋中030800 [2]重庆邮电大学理学院,重庆400065

出　　处：《河北师范大学学报(自然科学版)》2025年第3期217-233,共17页Journal of Hebei Normal University:Natural Science

基　　金：国家自然科学基金(12061067,62176033);重庆市自然科学基金(CSTB2023NSCQ-MSX0707)。

摘　　要：粒计算作为一种新的信息处理方法,能够发现数据中隐藏的一些多层次和多视角的知识,在众多领域有着广泛应用.基于粒计算理论提出的粒球计算方法,是以粒球的纯度为标准使用层次聚类进行信息粒化,并把粒球作为后续运算输入的方法.然而,由于粒球生成的逐步划分方式,导致了粒球生成结果的不稳定和多次迭代的高时间成本.本文引入了自然邻居聚类思想,对粒球生成展开了研究,提出了一种基于自然邻居局部中心搜索的自适应粒球生成算法.该算法得益于自然邻居局部中心搜索算法无参自适应的特点,对参数纯度阈值不敏感,因此是一种完全无参自适应的粒球生成算法.同时,由于本算法是对整体数据集进行搜索划分,因此每次能够生成更多的粒球并且减少迭代的次数.经实验验证,该算法在高纯度阈值或者高噪声率时相比GBG算法有速度优势和更高的稳定性,并且有着可以媲美GBG算法和KNN的分类精度.Granular computing,as a novel information processing method,can uncover multi-level and multi-perspective knowledge hidden within data,and has found extensive applications in various fields.Based on granular computing theory,the granular ball computing method utilizes hierarchical clustering to granulate information with the purity of granular balls as the criterion,using these granular balls as inputs for subsequent computations.However,the stepwise partitioning method of generating granular balls leads to instability in the results and a high time cost due to multiple iterations.This paper introduces the concept of natural neighbor clustering to study the generation of granular balls and proposes an adaptive granular ball generation algorithm based on local center search among natural neighbors.Benefiting from the parameter-free adaptiveness of the natural neighbor local center search algorithm,this method is insensitive to purity threshold parameters,making it a fully parameter-free,adaptive granular ball generation algorithm.Additionally,since this algorithm partitions the entire dataset in each iteration,it can generate more granular balls and reduce the number of iterations.Experimental validation shows that this algorithm offers speed advantages and higher stability under high purity thresholds or high noise levels compared to the granular ball generation algorithm,and it achieves classification accuracy comparable to that of the granular ball generation algorithm and KNN.

关 键 词：粒计算 自然邻居 粒球计算 信息粒 

分 类 号：Q938.8[生物学—微生物学]