Multiple Random Empirical Kernel Learning with Margin Reinforcement for imbalance problems

摘要

Imbalance problems arise in real-world applications when the number of negative samples far exceeds the number of positive samples, such as medical data. When solving the classification of imbalance problems, the samples located near the decision hyperplane contribute more to the decision hyperplane, and the samples far from the decision hyperplane contribute less to the decision hyperplane. So we can consider giving higher weights to the samples near the decision hyperplane, but they are sensitive to noise, and too much emphasis on them may lead to unstable performance. This paper proposes a Margin Reinforcement (MR) method to overcome the above dilemma. Because the imbalance problem is a cost-sensitive problem, MR gives positive samples a uniform high weight to improve the misclassification cost of the positive sample. For negative samples, according to their entropy, samples away from the decision surface and noise samples mixed in the positive samples are given a smaller weight, in order to improve the efficiency and robustness of the algorithm. Therefore, MR can emphasize the importance of samples located in overlapping regions of positive and negative classes and ignore the effects of noise samples to produce superior performance. Multiple Random Empirical Kernel Learning (MREKL) has proven to be effective and efficient in dealing with balance problems. In order to improve the performance of MREKL on imbalanced datasets, MR is introduced into MREKL to propose a novel Multiple Random Empirical Kernel Learning with Margin Reinforcement (MREKL-MR). MREKL-MR efficiently map the samples into low-dimensional feature spaces, then utilizes the MR approach to reenforce the importance of margin samples and decrease the effects of noise samples. Experimental results on 28 imbalanced datasets indicate that MREKL-MR is superior to comparison algorithms. Finally, the effectiveness of MREKL-MR in dealing with imbalance problems is verified on the Heart Failure dataset.