A Novel Ensemble Learning Method for Online Learning Scenarios

Abstract

Online Sequential Extreme Learning Machine (OS-ELM) is a modeling algorithm for data stream mining that has attracted much attention in recent years. OS-ELM has extremely fast training speed and good generalization ability, however, because it uses random input weights for the model initialization and these parameters remain unchanged throughout the training process, which causes instability in the performance of the model. To alleviate this problem, we propose an improved OS-ELM algorithm based on the ensemble learning mechanism, which we call ELOS-ELM. ELOS-ELM uses multiple initialization methods to construct diverse base online learners, and the prediction results of the model are determined by voting by these base models. In this way, the final model can avoid misjudgments caused by the knowledge blind zone of a single base model. Extensive experimental results on eight benchmarks show that ELOS-ELM can effectively improve the stability and generalization ability of the OS-ELM.