基于排序互信息的无监督特征选择

Unsupervised feature selection based on ranking mutual information

根据排序问题的单调先验知识,无监督学习问题中的观测属性之间也具备单调关系;否则该属性与排序无关,为冗余属性.基于排序互信息反应的两属性之间的单调关系,提出用每个属性与其他属性之间的平均互信息,来衡量每个属性与排序学习的相关程度,具有最高的平均互信息即为排序最相关的属性.

Based on ranking prior knowledge of monotonicity,each observation attribute should be monotonic with the other observation attributes for unsupervised ranking problems.Otherwise,the attribute would be irrelevant with ranking and should be assumed to a redundant attribute.Based on the ranking mutual information,which reflects the monotonic degree between observation attributes and the order sequence,mean ranking mutual information is proposed to measure the monotonicity between observation attributes.The most relevant attributes should be with the biggest ranking mutual information.