MinSVM for imbalanced datasets with a case study on Arabic comics classification -

Abstract

Class imbalance occurs when the different classification categories, or samples, are not equally represented in the training dataset. Class imbalance is frequent in many real life applications and particularly in Arabic short text classification. Classifying an imbalanced dataset is problematic because most traditional clas-sifiers achieve a high accuracy for the majority class, but a consistently low accuracy on the minority class. The many studies developed to classify standard Arabic text docu-ments do not perform well on Arabic short text due to the sparsity of the feature vector. This study proposes the Minority Support Vector Machines (MinSVM) classi-fier, a novel classifier based on Support Vector Machine for binary classification, a Root based Feature Reduction (RFR) scheme for short Arabic text. To validate the performance of our research, MinSVM was tested on some benchmark imbalanced datasets and on a Arabic comics datasets that was manually con-structed. In all our experiments, MinSVM results outperformed some of the main meth-ods suggested in literature for imbalance datasets