Home // International Journal On Advances in Software, volume 9, numbers 3 and 4, 2016 // View article

Semi-Supervised Ensemble Learning in the Framework of Data 1-D Representations with Label Boosting

Authors:
Jianzhong Wang
Huiwu Luo
Yuan Yan Tang

Keywords: Data smooth sorting; one-dimensional embedding; regularization; label boosting; ensemble classification; semi-supervised learning.

Abstract:
The paper introduces a novel ensemble method for semi-supervised learning. The method integrates the regularized classifier based on data 1-D representation and label boosting in a serial ensemble. In each stage, the data set is first smoothly sorted and represented as a 1-D stack, which preserves the data local similarity. Then, based on these stacks, an ensemble labeler is constructed by several 1-D regularized weak classifiers. The 1-D ensemble labeler extracts a newborn labeled subset from the unlabeled set. United with this subset, the original labeled set is boosted and the enlarged labeled set is utilized into the next semi-supervised learning stage. The boosting process is not stopped until the enlarged labeled set reaches a certain size. Finally, a 1-D ensemble labeler is applied again to construct the final classifier, which labels all unlabeled samples in the data set. Taking the advantage of ensemble, the method avoids the kernel trick that is the core in many current popular semi-supervised learning methods such as Transductive Supported Vector Machine and Semi-Supervised Manifold Learning. Because the proposed algorithm only employs relatively simple semi-supervised 1-D classifiers, it is stable, effective, and applicable to data sets of various types. The validity and effectiveness of the method are confirmed by the experiments on data sets of different types, such as handwritten digits and hyperspectral images. Comparing to several other popular semi-supervised learning methods, the results of the proposed one are very promising and superior to others.

Pages: 359 to 372

Copyright: Copyright (c) to authors, 2016. Used with permission.

Publication date: December 31, 2016

Published in: journal

ISSN: 1942-2628