Gravitation balanced multiple kernel learning for imbalanced classification

Mengping Yang, Zhe Wang, Yanqiong Li, Yangming Zhou, Dongdong Li, Wenli Du

Abstract: This paper considers a category of classification problems where the samples of different classes are not represented equally. It arises in a variety of application areas and has been widely studied in pattern recognition. This paper focuses on enhancing the original data representation by combining the gravitation-based method with multiple empirical kernel approach, this paper proposes a sample level method known as the gravitational balanced multiple kernel learning (GBMKL) method. Our proposed GBMKL method integrates gravity strategy to generate the gravitation balanced midpoint samples (GBMS) located on the classification boundary; meanwhile, the classification boundary can be rectified by the nearest neighbors of the boundary (NNB) samples, which can improve the generalization performance. We further design two regularization terms corresponding to GBMS and NNB to avoid overfitting. In the training and testing process, the samples are mapped into multiple empirical kernel spaces to obtain more sufficient data representation. We conduct extensive computational experiments on 54 imbalanced datasets including both artificial and real-word datasets selected from knowledge extraction based on evolutionary learning datasets, the experimental results reveal interesting insights and show the advantages of the proposed GBMKL approach for dealing with the imbalanced classification problems. In addition, parameter analysis of two regularization terms confirms their positive impacts on the classification performance.