Semi-supervised Learning from Unbalanced Labeled Data – An Improvement

Abstract

We present a possibly great improvement while performing semi-supervised learning tasks from training data sets when only a small fraction of the data pairs is labeled. In particular, we propose a novel decision strategy based on normalized model outputs. The paper compares performances of two popular semi-supervised approaches (Consistency Method and Harmonic Gaussian Model) on the unbalanced and balanced labeled data by using normalization of the models’ outputs and without it. Experiments on text categorization problems suggest significant improvements in classification performances for models that use normalized outputs as a basis for final decision.