Towards Ensemble-Based Imbalanced Text Classification Using Metric Learning

Abstract

This paper reports s series of ensemble approaches for imbalance text classification. All the approaches utilize a metric learning technique for obtaining better representations of texts to train weak classifiers. Each approach deals with the class imbalance problem with an undersampling-based ensemble approach, because metric learning techniques also suffer from this problem. In this paper, four ensemble approaches (namely, MLBagging, MLBoosting, MLStacking, and MLBoostacking) are proposed, three of which are corresponding to ensemble frameworks (namely, bagging, boosting, and stacking), and the other is a combination of boosting and stacking. MLBagging, MLBoosting, and MLStacking train metric learners on the individual undersampled dataset and combine them, while MLBoostacking trains metric learners in a step-by-step manner; that is, a metric learner learns a feature transformation so that failed-to-classify samples in the previous step should be correctly classified. The experimental evaluation on three imbalanced text classification datasets (namely, unfair statement classification in terms of service, hate speech detection in a forum, and hate speech tweet detection) shows that the proposed approaches lift classification performance from BERT-based approaches, by improving the representations of texts through metric learning.