Active learning by increasing model likelihood for Gaussian mixture models based classifiers

• Published in: IOP conference series. Materials Science and Engineering 2021-03, Vol.1098 (3), p.32036
• Main Author: Iswanto, B H
• Format: Article
• Language: English
• Subjects: Active learning; Algorithms; Classifiers; Machine learning; Optimization; Probabilistic models
• ISSN: 1757-8981; 1757-899X
• DOI: 10.1088/1757-899X/1098/3/032036

A problem in many classification tasks is to acquire labelled data, while large amounts of unlabelled data are available. One way to overcome these problems is to apply active learning. This technique aims to select the most informative examples and to build optimally classifiers. In this paper, we propose an active learning algorithm for Gaussian mixture model classifiers by maximizing the current model likelihoods. The method assumes a large pool of unlabelled examples is available, the examples are i.i.d according to some underlying distribution, and the labels are distributed according to the class-conditional distribution. Experiments using artificial and real datasets show that, the proposed method outperform and show efficiency in the context of the query number compared with random and expected likelihood queries.