Variances-constrained weighted extreme learning machine for imbalanced classification

Abstract

Extreme learning machine (ELM) is a widely used algorithm due to its fast speed and good generalization. But ELM performs well under the assumption that the data is balanced. Because in binary classification with high imbalanced ratio data, the accuracy of samples from the minority class tends to be ignored easily due to the disadvantage in quantity. Therefore, weighted ELM (W-ELM) was proposed for imbalance learning. When samples are transformed through the hidden layer of W-ELM, W-ELM needs to calculate suitable weights to classify the samples linearly. But it is difficult for W-ELM to achieve a good learning performance when the dispersion of the transformed samples is high enough. Besides, W-ELM is limited by its two weighting schemes. Many W-ELM-based methods, such as boosting W-ELM, deep W-ELM and ensemble W-ELM, were also proposed with the limitations of W-ELM. Class-specific cost regulation ELM (CCR-ELM) is an alternative approach to learn imbalanced data. However, to a certain degree, CCR-ELM is similar to ELM. Therefore, in this paper, we will propose variances-constrained weighted extreme learning machine (VW-ELM) to replace W-ELM. VW-WLM takes the dispersion of training results into consideration. From theoretical view, VW-ELM’s learning performance is no lower than W-ELM. We use 21 benchmark datasets to test the learning ability of VW-ELM. Experimental results demonstrate that VW-ELM has an obvious advantage to handle classification problem of imbalanced data, comparing with the state-of-the-arts.