Adaptive Resampling with Active Learning

This paper proposes a novel algorithm Virtual Instances Resampling Technique Using Active Learning (VIRTUAL) for class imbalance problem in Support Vector Machine (SVM) learning. In supervised learning, prediction performance of the classification algorithms deteriorate when the training set is imbalanced. Class imbalance problem occurs when at least one of the classes are represented by substantially less number of instances than the others in the training set. Various real-world classification tasks, such as medical diagnosis and text categorization suffer from this phenomenon. VIRTUAL is a hybrid method of oversampling and active learning to form an adaptive technique for resampling of the minority class instances. Unlike traditional resampling methods which require preprocessing of the data, VIRTUAL generates virtual instances for the minority class support vectors during the training process, therefore it removes the need for an extra preprocessing stage. Our empirical results show that VIRTUAL outperforms other competitive oversampling techniques and active learning strategy in terms of prediction capability. In addition, VIRTUAL is more efficient in generating new instances and has a shorter training time than the other oversampling techniques due to its adaptive nature and its decision capability in creating virtual instances.