Semi-supervised text classification: Partial training from unlabeled data

We illustrate by a case study how a semi-supervised approach can improve the performance of text classification. We begin with a naïve Bayes classifier trained exclusively from labeled text documents, and apply it to a set of unlabeled text documents to derive their pseudo-labels. The pseudo-labels are then combined with the true labels in the original training sample, and a naïve Bayes classifier is built based on the enlarged training sample. We consider different proportions of pseudo-labels in the enlarged training sample, and examine the effect of the semi-supervised approach on the misclassification rate by using cross validation comparisons.