Ensemble online weighted sequential extreme learning machine for class imbalanced data streams

Abstract

Class imbalanced data streams often have unbalanced sample distribution and a certain type of sample size is too small, which will lead to over fitting phenomenon due to insufficient sample learning, and most current classifiers have problems such as model instability. Therefore, choosing the online sequential extreme learning machine (OSELM) as the basic theoretical model, and combining the AdaBoost ensemble learning ideas and cost-sensitive strategies, an ensemble online weighted sequential extreme learning machine algorithm (ABC-OSELM) is proposed. Firstly, in order to solve the problem that the minority classes are easily misclassified due to class imbalance, the OSELM algorithm based on cost-sensitive learning (C-OSELM) is proposed, which improves the misclassification by assigning different penalty parameters to various samples, it can effectively alleviate the phenomenon of excessive deviation of the decision-making surface. On this basis, in order to further improve the classification accuracy and stability of the algorithm, combining C-OSELM with ensemble learning ideas, an ensemble C-OSELM algorithm based on AdaBoost (ABC-OSELM) is proposed. By adopting a homogeneous integration strategy, iteratively adjust the weight of the base classifier to generate a more stable strong classifier. Finally, the effectiveness and feasibility of the ABC-OSELM algorithm are verified through 15 class II imbalanced datasets.