A Selective Under-Sampling based Bagging SVM for Imbalanced Data Learning in Biomedical Event Trigger Recognition

Abstract

Because of the importance of the biomedical events in understanding the biomedical processes and functions, extraction of events from biomedical literature automatically becomes the recent research focus, which can update the existing biomedical knowledge faster. Recognition of trigger words is a very important first step for event extraction, since the following extraction steps depends on the outputting trigger words. The application of event trigger recognition aims to classify different kinds of trigger words from very large biomedical literature data sets with significant imbalance between classes. However, many existing classification methods show great limitations and performance decline on imbalanced and large data sets. In this paper, a selective under-sampling based bagging Support Vector Machine, SUS Bagging-SVM, approach has been proposed to address the issues together in an ensemble learning framework. Through making use of boundary information, a novel selective under-sampling method was designed to prune unimportant instances from the majority class to reduce imbalance in large data sets effectively. SUS Bagging-SVM was tested on the corpus of BioNLP'09 shared task. It achieved a total F1-measure of 66.1, which is competitive in comparison with other state-of-art trigger recognition systems, and a higher recall value. In conclusion, SUS Bagging-SVM is a valuable method for alleviating the problem of classification on imbalanced and large data.