This study extends the previously reported feature assessment scheme CORPS (class overlap region partitioning scheme) designed for perfectly supervised environments to the imperfectly supervised domain. The imperfectness levels of the labels, which can be different for different classes, are used to appropriately weight the feature space overlap evaluation process, i.e., the samples from classes with more reliable labels are given correspondingly more weightage than those with less reliable labels. The methodology can be applied to mixed supervised and imperfectly supervised environments also, with subsets of data even within a class having different imperfectness levels. Like CORPS, the extended method can be used either as a stand alone tool or as a front end to more complex combinatorial feature selection procedures such as branch and bound and genetic algorithms. The new approach also has the flexibility to permit a bias in favor of either of the two possible types of errors in a binary decision process, such as false alarm and leakage in a target detection problem. Algorithmic and operational details are included to facilitate wide usage of this new tool.
展开▼
