一种基于主成分分析的元学习方法

A Meta Learning Method Based on Principal Component Analysis

Meta-learning是解决分布式环境下分类挖掘问题的有效方法之一,该方法的关键在于元训练特征的组成,它直接影响着元分类器的性能.当问题域中类别数较多时,组合多个基分类器将使元训练特征集的维度变得非常大.提出在元学习阶段用主成分分析(PCA)简化高维数据,将基分类器的预测信息集中到最小维数的向量上,以提高元学习的泛化效率.实验结果表明,该方法是可行的.

Meta-learning is one of the effective classifier combination techniques,which is used for distributed mining of classification.The key problem of this method is that the performance of meta-classifier depending on the feature vector of meta training set.Combining large quantities of base classifiers for domains with many classes will produce high-dimension feature vector.This paper presents a method that employs Principle Component Analysis（PCA） to extract high dimension data for meta learning,then the predictive information of base classifiers will concentrate on the minimum dimension of the vector,which will improve the generalization efficiency for meta learning.Experimental evaluation shows that the new algorithm is feasible and efficient.