基于分类编码的集成学习结构多样性研究  

Research on structural diversity of ensemble learning based on classification coding

作　　者：周钢[1] 崔良中[1] ZHOU Gang;CUI Liang-zhong(College of Electronic Engineering, Naval Univ. of Engineering, Wuhan 430033, China)

机构地区：[1]海军工程大学电子工程学院,武汉430033

出　　处：《海军工程大学学报》2021年第2期78-83,90,共7页Journal of Naval University of Engineering

摘　　要：针对传统结构多样性方法的度量不精准和在异质基学习器不通用的问题,提出了分类编码多样性方法。该方法首先根据基学习器在数据集分块上的表现构建三值分类编码,以基学习器分类编码的绝对差值来度量成对多样性,并构建全局多样性度量最优化解;然后,按照贪婪算法完成基学习器的选择性集成;最后,在分类数据集上对比普通Bagging算法和树结构匹配算法后发现,分类编码多样性方法比树结构匹配算法具有更好的适用性、解释性和计算性,比普通Bagging算法具有更高的泛化精度,特别是在二分类问题上表现更优。The classification coding diversity(CCD)method was proposed to solve the problems of inaccurate measurement of traditional structural diversity methods and incomplete use of heterogeneous base learners.The method constructed three valued classification coding according to the performance of the basic learning device in the data block partition,then the absolute difference of those was used to measure the pairwise diversity,and the global diversity measure was used to construct the most optimal solution.The greedy algorithm was used to complete the selective integration of the basic lear-ning device.Compared with ordinary Bagging algorithm and tree structure matching(TMD),the method of classification coding diversity(CCD)has better applicability,explanatory and computational performance than the tree structure matching(TMD)algorithm,and has higher generalization accuracy than the Bagging algorithm,especially for the two-classification problem.

关 键 词：集成学习 结构多样性 分类编码 贪婪算法 BAGGING 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]