摘要
由于类分布的不平衡性,很多传统的分类方法在非平衡数据集上的分类效果不好.与传统的方法不同,论文从组合选择的角度考虑不平衡类学习问题,提出了一种基于实例的组合选择方法 CBES,提升组合分类器在不平衡数据集上的分类性能.CBES考察类标号未知的样本的k近邻,并以此为选择集,从分类器库中选择一个最优或次优的子组合分类器来预测未知样本的类标号.由于考虑了待分类样本的局部特征,更关注稀有类,因此CBES能够更好地对非平衡数据集进行分类.实验结果表明,本文的方法能够显著地降低模型的复杂度,有效地提高了非平衡数据集上的分类性能.
Because of imbalanced class distribution, most classifiers do not work well on imbalanced datasets. Unlike traditional meth- ods, this paper reconsiders class imbalance learning problem from the viewpoint of ensemble selection, and proposes a new method, named case-based ensemble selection approach (CBES), to improve classifiers' performance on these datasets. CBES tries to search for the k-nearest neighbors of an unlabeled instance as the selection set, and then select an optimal or suboptimal sub-ensemble to pre- dict the instance. Due to considering the local properties of the unlabeled instance and taking more attention on rare class, CBES can do classification more efficiently on imbalanced data sets. Experimental results indicate that CBES induces models significantly with better performance on almost all imbalanced datasets comparing to traditional ensemble approaches.
出处
《小型微型计算机系统》
CSCD
北大核心
2014年第4期770-775,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(60773048
61170223)资助