Probability cost function based weighted extreme learning machine

Standard extreme learning machine has good generalization performance and fast learning speed, but has the disadvantage of degrading performance for imbalance learning. Weighted extreme learning machine (WELM) is a kind of cost-sensitive learning method that significantly improves the classification performance for imbalanced data by adding extra weight to each training sample. In this paper, we present a novel WELM that defined by probability cost function concerned with the probability that given sample belong to each class. We propose learning network mapping input training data to a vector consisting of the probability value of given training sample belonging to each class. We define its cost functions to maximize the marginal distance between classes and the probability that each training sample will be accurately classified. We empirically show that our proposed algorithm obtains superior performance in general than some state-of-the-art imbalance learning approaches on 32 binary class and 14 multiclass imbalanced datasets. To further estimate the experimental results, we also provide statistical analysis.