Evolutionary extreme learning machine with sparse cost matrix for imbalanced learning.

Extreme learning machine is a popular machine learning technique for single hidden layer feed-forward neural network. However, due to the assumption of equal misclassification cost, the conventional extreme learning machine fails to properly learn the characteristics of the data with skewed category distribution. In this paper, to enhance the representation of few-shot cases, we break down that assumption by assigning penalty factors to different classes, and minimizing the cumulative classification cost. To this end, a case-weighting extreme learning machine is developed on a sparse cost matrix with a diagonal form. To be more actionable, we formulate a multi-objective optimization with respect to penalty factors, and optimize this problem using an evolutionary algorithm combined with an error bound model. By doing so, this proposed method is developed into an adaptive cost-sensitive learning, which is guided by the relation between the generalization ability and the case-weighting factors. In a broad experimental study, our method achieves competitive results on benchmark and real-world datasets for software bug reports identification.