A multi-view approach to semi-supervised document classification with incremental Naive Bayes

Abstract

Many semi-supervised learning algorithms only consider the distribution of word frequency, ignoring the semantic and syntactic information underlying the documents. In this paper, we present a new multi-view approach for semi-supervised document classification by incorporating both semantic and syntactic information. For this purpose, a co-training style algorithm, Co-features, is proposed. In the phase of active querying, we assign a weight to each sample document according to its uncertainty factor. Then the most informative samples are selected and labeled by other “teachers”. In contrast to batch training mode, we developed an incremental Naive Bayes update method, which allows for more efficient training even with a large pool of unlabeled data. Experimental results show that our algorithm works successfully on the datasets Reuters-21578 and WebKB, and is superior to Co-testing in the learning efficiency.