An ensemble approach to handle out of vocabulary in multilabel document classification

Abstract

In document classification, out of vocabulary occured when there are words that are not covered by the classifier. This condition can affect prediction performance. The classifier needs to be re-trained using new documents containing those words in order to handle out of vocabulary condition. In this paper, we proposed an ensemble method that able to re-train a classifier in incremental fashion. That is, it does not use any previous training documents, only the new ones. Our approach also handles multilabel classification where a document can be classified into more than one label, which is a the case to Indonesian news articles we use for this paper. Result from our experiment shows that the ensemble method can cover out of vocabulary words and correctly classify a document where other classifiers which do not cover the words can not. It some conditions, it can even outperform batch-trained classifier performance. And also, because our ensemble method is an incremental method, its training time and memory consumption depends to the training data it currently uses.