最大化最小margin的抽样多样性集成学习方法研究  

SAMPLING DIVERSITY METHOD IN ENSEMBLE LEARNING WITH MAXIMIZE THE MINIMUM MARGIN

作　　者：周钢 郭福亮[1] Zhou Gang;Guo Fuliang(Naval University of Engineering,Wuhan 430033,Hubei,China)

机构地区：[1]海军工程大学,湖北武汉430033

出　　处：《计算机应用与软件》2020年第8期300-306,共7页Computer Applications and Software

摘　　要：从偏差-方差分解和误差-分歧分解角度,发现集成学习中的基分类器多样性是提升集成学习泛化能力和分类精度的重要途径,在数据样本中通过最大化最小margin的方法能够有效提升集成学习多样性。针对Bagging算法,提出一种基于最大化最小margin的优化集成学习算法。通过优化抽样方法实现最大化最小margin,将该问题简化为基分类器权重调整问题;通过给正确分类易错数据样本的基分类器赋予更高权重来实现在集成学习抽样中该类数据样本的margin值最大化。实验结果表明,该算法对比常用集成学习方法在大部分数据集上具有更高的执行效率和泛化精度。From the perspective of deviation-variance decomposition and error-bifurcation decomposition,it is found that the diversity of base classifiers in ensemble learning is an important factor to improve the generalization ability and classification accuracy of ensemble learning.The method of maximizing the minimum margin in the data samples can effectively improve the diversity of ensemble learning.For Bagging algorithm,we propose an optimal ensemble learning algorithm based on the minimum margin maximization.It optimized the sampling method to maximize the minimum margin,and simplified the problem to the weight adjustment of the base classifier.By assigning a higher weight to the base classifier of the correct classification of error-prone data sample,the margin value of this kind of data samples in integrated learning sampling was maximized.The experimental results show that our algorithm has higher execution efficiency and generalization accuracy than the commonly used ensemble learning methods on most data sets.

关 键 词：集成学习 多样性 MARGIN BAGGING 动态权重 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]