Improving SMOTE with Fuzzy Rough Prototype Selection to Detect Noise in Imbalanced Classification Data

In this paper, we present a prototype selection technique for imbalanced data, Fuzzy Rough Imbalanced Prototype Selection (FRIPS), to improve the quality of the artificial instances generated by the Synthetic Minority Over-sampling TEchnique (SMOTE). Using fuzzy rough set theory, the noise level of each instance is measured, and instances for which the noise level exceeds a certain threshold level are deleted. The threshold is determined using a wrapper approach that evaluates the training Area Under the Curve of candidate subsets. This proposal aims to clean noisy data before applying SMOTE, such that SMOTE can generate high quality artificial data. Experiments on artificial data show that FRIPS in combination with SMOTE outperforms state-of-the-art methods, and that it particularly performs well in the presence of noise.