An Imbalanced Data Classification Method Based on Multi-Kernel Extreme Learning Machine Fusion

To solve the problem of low classification accuracy in imbalanced data classification, a method based on Multi-Kernel Extreme Learning Machine Fusion (MKELMF) is proposed. By introducing the “learning to learn” idea of meta-learning algorithm, the ability of Multi-Kernel Extreme Learning Machine (MK-ELM) to classify local areas of data samples is fully mined. For each test sample, the meta-learning algorithm selects a group of Kernel Extreme Learning Machine (KELM) with sufficient classification ability for combined decision making, avoiding the problems such as insufficient classification of single KELM and the problem that ordinary MK-ELM can easily introduce wrong classification information by making weighted decisions based on all KELM results. According to the characteristics of KELM, five evaluation criteria are proposed as meta-features for training meta-learner. We construct a serial regulating circuit and collect data under normal and fault conditions to verify the proposed method. The experimental results show that compared with two cost-sensitive SVM methods and two local multi-kernel weighted fusion methods, the MKELMF method improves the test accuracy, Macro-F1 value and G-mean by 8.53%, 4.59% and 5.37% respectively. In addition, the method has higher training and testing efficiency.