K Nearest Gaussian-A model fusion based framework for imbalanced classification with noisy dataset

Abstract

Data quality issues such as data imbalance and data noise have great impact on the performances of many classifiers. Althoughthe co-existence of imbalance and noise appears in many real world datasets, the issue of imbalance and noise have mostlybeen treated separately due to their different causes and problematic consequences. However, doing so may ignore the mutualeffects thus may not achieve optimal classification performance. In this research, we propose a model fusion based framework,termed K Nearest Gaussian (KNG) to tackle the imbalance and noise issues jointly. KNG employs generative modeling method(GMM) to extract the data characteristics from the training data which are less sensitive to data imbalance and noise. The datacharacteristics are then used to establish Gaussian confidence regions which are used to achieve final classification in a K nearestneighbor (KNN) manner. Experiments on seven UCI benchmark datasets and one medical imaging dataset show KNG methodgreatly outperforms traditional classification methods in dealing with imbalanced classification problems with noisy dataset.