The effect of unlabeled data on generative classifiers, with application to model selection
Ira L. Cohen, Fábio Gagliardi Cozman, Alexandre Bronstein · 2002
semisupervised learning, labeled and unlabeled data problem, classification, machine learning In this paper we investigate the effect of unlabeled data on generative classifiers in semi-supervised learning. We first characterize situations where unlabeled data cannot change estimates obtained with labeled data, and argue that such situations are unusual in practice. We then report on a large set of experiments involving labeled and unlabeled data, and demonstrate that unlabeled data can degrade classification performance when modeling assumptions are incorrect. To improve classification performance, we propose a method to switch assumed model structure based on the effect of unlabeled data.