准确度优先的多目标帝王蝶优化定量构效特征选择方法  

Accuracy Priority Based Multi-objective Monarch Butterfly Optimization for Feature Selection in Quantitative Structure-activity Relationship

作　　者：李小林[1] 王静 张元孜 黄世国[1] LI Xiao-lin;WANG Jing;ZHANG Yuan-zi;HUANG Shi-guo(Computer and Information College,Fujian Agriculture and Forestry University,Fuzhou 350002,China)

机构地区：[1]福建农林大学计算机与信息学院,福州350002

出　　处：《小型微型计算机系统》2021年第12期2506-2510,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(31870641)资助;福建省自然科学基金项目(2017J01607,2018J01612)资助;福建农林大学科技创新专项基金项目(KFA17030A,KFA17181A)资助;福建省林业科技项目(闽林科便函[2018]26号)资助。

摘　　要：针对帝王蝶优化算法用于特征选择时需满足多目标的要求,对该算法进行了3个方面的改进:(1)在个体排序步骤中引入非支配排序算法,并对调整算子做了修正,满足了多目标要求;(2)增加了准确度优先策略,减少了计算资源在低准确性区域的搜索,保证了模型的准确性,满足了特征选择中准确性优先于特征数的要求;(3)增加了基于子组的突变策略,对不同子组使用不同的突变策略,避免了算法过早陷入局部最优,解决了算法早熟问题.在3个定量构效特征选择基准数据集上进行了一系列实验,实验结果表明改进的算法与其它算法相比显著提高了模型的准确性并减少了特征数,证明了改进策略的有效性.In order to meet the requirement of multi-objective in feature selection,this paper improves Monarch Butterfly Optimization algorithm in three aspects: (1) Non-dominated Sorting Algorithm is applied to the individual sorting step,and the adjusting operator is modified to meet the multi-objective requirements. (2) The accuracy priority strategy is utilized to reduce the computing resources of searching in the low accuracy area,which ensures the accuracy of the model and meets the requirement that accuracy takes precedence over the number of features in feature selection. (3) The different mutation strategies based on different subgroups are introduced to avoid trapping into local optimum,and solving the problem of premature convergence. A series of experiments were carried out on three benchmark data sets for feature selection in Quantitative Structure-Activity Relationship modeling. The experimental results showed that the improved algorithm significantly improves the accuracy of the model and reduces the number of features compared with other algorithms,which proves the effectiveness of the proposed strategies.

关 键 词：特征选择 帝王蝶优化 非支配排序 定量构效 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]