A new over-sampling technique based on SVM for imbalanced diseases data
Jinjin Wang, Yukai Yao, Hanhai Zhou, Mingwei Leng, Xiao Yun Chen · 2013
In the real world, there are many kinds of diseases data, whose patients are composed of majority normal persons and only minority abnormal ones. Many researchers ignored these imbalance problems, so their learning models usually led to a bias in the majority normal class. To deal with this problem, a new over-sampling technique was proposed to over-sample the minority class to balance the data samples and improve Support Vector Machine(SVM) in imbalanced diseases data sets. For the minority class, a K-Nearest Neighbor(KNN) graph is built. Second, the proposed technique gets a Minimum Spanning Tree(MST) based on the graph. Third, the proposed technique generates synthetic samples by using SMOTE along the direct path in the tree. The performance of the proposed technique based on SVM is evaluated with several diseases data sets taken from the UCI machine learning repository, and the experiments show that the proposed technique based on SVM can improve the Sensitivity value and G-Mean value.