A0258
Title: A greedy active learning algorithm in multinomial logistic regression
Authors: Hsiang-Ling Hsu - National University of Kaohsiung (Taiwan) [presenting]
Abstract: For building a proper classification system to predict the class type of data, large amounts of labeled training samples are needed, which might result in lots of resources to obtain the effective labeled information. For this issue, we can adopt active learning to recruit the crucial data from a massive unlabeled data set, and then obtain its labeled information, finally put it into a labeled data set, which is utilized to construct the classifier. For a binary data analysis with the logistic regression models, the GATE algorithm not only considers the subject selection scheme but also integrates the variable extraction step to build the classifier more efficiently. An active learning procedure of multiple-class classification data has been previously proposed via individualized binary models for both categorical and ordinal labeled data. Moreover, for the active learning, it has been shown that selecting the initial samples effectively assists in building the classification model. We construct an active learning procedure that integrates the concepts of the initial samples determination, subject screening and variable selection simultaneously for applying multiple-class classification problems. Simulation studies and the analyzed results of open data sets demonstrate the classification performances for the proposed algorithm.
