New approaches to imbalanced text classification

The distribution of text data across classes is often imbalanced. This condition leads to classifiers tending to perform poorly on smaller categories within imbalanced data sets. As a result, text classification is a process significantly affected by the imbalanced class problem. The feature selection stage, one of the crucial stages of the text classification process, is also important for the imbalanced text classification problem. In this thesis, the problems of feature selection for text classification and the solutions offered by popular feature selection methods are extensively analyzed, and various solutions are proposed for the feature selection stage. To this end, firstly, the effect of feature selection methods on the classification of imbalanced texts is thoroughly examined. In this direction, many experiments were carried out with three different classifiers and nine different feature selection methods on two different data sets. Additionally, the success of feature selection methods has been observed using different numbers of features. Also, two new feature selection methods (EFS_IMP1 and EFS_IMP2) were proposed for imbalanced text classification. These methods are derived from a recent feature selection method called Extensive Feature Selector (EFS). The performance comparison of EFS_IMP1 and EFS_IMP2 methods was carried out with six filter-based feature selection methods. Three benchmark imbalanced text data sets were employed with Support Vector Machines (SVM), Decision Tree (DT), Random Forest (RF), and K-Nearest Neighbors (kNN) classifiers. Experimental results showed that EFS_IMP1 and EFS_IMP2 offer superior or comparative performance compared with other feature selection methods based on Macro-F1 for imbalanced text classification.