A synthetic informative minority over-sampling (SIMO) algorithm leveraging support vector machine to enhance learning from imbalanced datasets

Developing decision support systems (DSS) based on imbalanced datasets is one the critical challenges in data mining and decision-analytics. A dataset is called imbalanced when the number of examples from one class outnumbers the number of the instances from another class. Learning from imbalanced datasets is one of the major challenges in machine learning. While a standard classifier could have a very good performance on a balanced dataset, when applied to an imbalanced dataset, its performance deteriorates dramatically. This poor performance is rather troublesome, especially in detecting the minority class, which usually is the class of interest. Therefore, the poor performance of machine learning techniques, which are used to develop DSS, negatively affect the practicality of DSS in real word problems. Over-sampling the minority class is one of the most promising remedies for imbalanced data learning. In this study, we propose a new synthetic informative minority over-sampling (SIMO) algorithm leveraging support vector machine (SVM). In this algorithm, first SVM is applied to the original imbalanced dataset, then, minority examples close to the SVM decision boundary, as the informative minority examples are over-sampled. We also developed another version of SIMO and call it weighted SIMO (W-SIMO). W-SIMO is different from SIMO in the degree of over-sampling the informative minority examples. In W-SIMO, incorrectly classified informative minority examples are over-sampled with a higher degree compared to the correctly classified informative minority examples. In this way, there is more focus on incorrectly classified minority examples. The over-sampled dataset can be used to train any classifier. We applied these algorithms to the 15 publicly available benchmark imbalanced datasets and assessed their performance in comparison with existing approaches in the area of imbalanced data learning. The results showed that our algorithms had the best performance in all datasets compared to other approaches.