不平衡超限学习机的全局惩罚参数选择方法

摘要

Conventional extreme learning machines (ELMs) usually perform poorly in learning and classifying im-balanced datasets, because positive samples are likely to be misclassified.However, weighted extreme learning ma-chine only considered between-class imbalance but ignored within-class imbalance.This paper explained why ELMs failed, and proposed a direct method to determine the penalty parameter, we considered both of the two kinds of imbalance, combine the between-class cost parameter with within-class cost parameter to form global penalty param-eter, that was, class penalty parameterwas refined further to single sample cost parameter.Theory analysis and sim-ulation experiments showed that the global penalty parameter selection for extreme learning machine is convenient in implementation, and performed better in improving the classification accuracy than some other types of extreme learning machine.%超限学习机在对不平衡数据集进行学习和分类时,正类样本容易被错分.而加权超限学习机只考虑了数据集类之间的不平衡,忽视了样本类内的不平衡的现象.本文阐述了超限学习机在不平衡数据集上分类效果欠佳的原因,提出了根据数据集选取惩罚参数的方法,采用将类间的惩罚参数与类内的惩罚参数相结合的方法,形成全局惩罚参数,即将类惩罚参数进一步精确到样本个体惩罚参数.结果表明:这种方法实现起来简单方便,与其他类型的超限学习机相比较,这种全局惩罚参数的选择方法在提高分类准确率方面能够取得更好的效果.