基于加权熵的重要性不对等样本学习的知识约简

由于学习样本的分布特性和所反映的主观特性的不同,每一个样本相对于学习算法的重要性程度往往是不对等的,为了能够在知识约简过程中考虑到学习样本的不对等性,提出了基于加权熵的知识约简方法。针对各类别样本分布不均匀的样本重要性不对等情况,为了加强小类样本所蕴含的知识在知识约简中的体现,提出一种逆类概率加权的不对等样本加权方法,实验表明该方法能够明显提高小类样本的分类正确率,也验证了基于加权熵的知识约简方法能够将样本的不对等性体现在知识约简结果中。

Weighed Entropy Based Knowledge Reduction in Learning from the Significance-imbalanced Instances

Since the difference of the distribution and subjective characteristic of the training instances,the significance of each instance is usually unequal in a learning algorithm.In order to take the imbalance of training instances into account,this paper proposes an approach to knowledge reduction based on weighted entropy.Aiming at the imbalance of class distribution,this paper presents an inverse class probability weighting approach to intensify the small class instances in knowledge reduction.The experiments express that the weighting approach obviously enhances the classification accuracy of the small class instances,which explains the validity of the approach to knowledge reduction based on weighted entropy in dealing with the imbalance of instances.