| 摘要: |
| 分析Stacking框架的基本原理,T1空间的数据表示和Stacking的缺陷,认为基于Stacking框架的学习能够有效地提高学习效果,但是在分类器个数增大时可能会导致元层训练数据规模增加。提出对底层分类器输出的后验概率用加权平均的方法构造元训练样本,减少二次建模的时间开销.该方法能够弥补由于对平均后验概率进行简单平均而丧失的模型输出特征,纠正分类偏差 |
| 关键词: 学习机制 Stacking 元学习 分类 分类器组合 |
| DOI： |
| 投稿时间：2004-05-17 |
| 基金项目:广西留学回国人员科学基金项目(桂科回0342001);广西教育厅科技项目(桂教科研[2001]401号)联合资助。 |
|
| A Researches on Learning Mechanism Based on Stacking Framework |
|
Wei Yanyan, Li Taoshen
|
| (Coll. of Comp. & Elec. Info., Guangxi Univ., Nanning, Guangxi, 530004, China) |
| Abstract: |
| One of disadvantages in Stacking method of the classifier is that the size of meta training examples increases when the number of base classifiers goes up.An approach to overcome this disadvantage is presented,in which the weighted average distribution of the posterior probability of the classifiers is used to form the meta training set.Experiment result shows that this approach can improve the output character lost caused by the average distribution of the posterior probability,and correct the mistakes made by base classifiers. |
| Key words: learning mechanism stacking meta learning classification classifiers combination |
