您的位置：首页 > 论文页面
Stacking 学习与一般集成方法的比较研究
发表时间：2018-02-28 浏览量：2283 下载量：838
| 全部作者： | 鲁莹，郑少智 |
| 作者单位： | 暨南大学经济学院 |
| 摘 要： | 集成学习（ensemble learning）因通过组合多个学习器实现更强的泛化能力而被广泛使用。目前一般的集成方法如AdaBoost、Bagging等均是基于一种算法，而Stacking集成是基于多种算法。本文针对分类集成问题，基于朴素贝叶斯（naive Bayes，NB）、logistic回归、k最近邻（k nearest neighbor，KNN）、决策树（C4.5）和规则学习（rule learner）5种基分类器，构建Stacking学习框架，并与AdaBoost、Bagging、随机森林（random forests，RF）、投票表决及交叉验证下的最佳分类器这5种方法进行比较。通过2组模拟数据和36组真实数据的实证分析发现，Stacking在所有方法中表现最好，具有最强的泛化能力且更适合大样本的情况。 |
| 关 键 词： | 人工智能；集成学习；组合；Stacking；分类器；泛化能力 |
| Title： | A comparative study of Stacking learning and general ensemble methods |
| Author： | LU Ying, ZHENG Shaozhi |
| Organization： | College of Economics, Jinan University |
| Abstract： | Ensemble learning has been widely used because of the great generalization ability by combining multiple learners. The general ensemble methods, such as AdaBoost and Bagging, are usually based on the same algorithm, but Stacking ensemble is based on different algorithms. Aiming at classification ensemble problems, this paper constructs a learning framework for Stacking consisted of 5 types of base classifiers, which are naive Bayes (NB), logistic regression, k nearest neighbor (KNN), decision tree and rule learner. And comparasion of Stacking with 5 other methods, AdaBoost, Bagging, random forests (RF), voting plan and selecting the optimal classifier using cross validation, is made. The experiments of 2 simulated and 36 real datasets are conducted and the results show that Stacking is the best among all other methods according to the highest generalization ability and more suitable for large sample cases. |
| Key words： | artificial intelligence; ensemble learning; combination; Stacking; classifier; generalization ability |
| 发表期数： | 2018年2月第4期 |
| 引用格式： | 鲁莹，郑少智. Stacking 学习与一般集成方法的比较研究[J]. 中国科技论文在线精品论文，2018，11（4）：372-379. |
0
评论数 0请您登录
暂无评论
下载全文
多维论文