Automatic filtering algorithm for imbalanced classification

The imbalanced data set has been reported to hinder the classification performance of many machine learning algorithms on both accuracy and speed. But extremely imbalanced data sets (3~5% positive samples) are common for many applications, such as multimedia semantic classification. In this paper, we propose a novel algorithm to automatically remove samples that have no or negative effects on classifier training for imbalanced training data sets. By using our algorithm, most easy-to-classify dominant-class samples in imbalanced training set will be eliminated automatically. As a result, the ratio of minority class samples is increased significantly, making it more suitable for classification algorithms. Experiments show that our algorithm can keep the classification accuracy of SVM, and decrease the training time dramatically.