A Semi-supervised Online Sequential Extreme Learning Machine Method

Abstract

Online sequential ELM (OS-ELM) provides a solution for streaming data application by only learning the newly arrived single or chunk of observations, and presents outstanding performance for learning problems. However, the algorithm relies on the labeled data, which usually involves high cost in labor and time. Moreover, manually labeled data suffers from inaccuracy caused by individual bias. Considering the semi-supervised ELM (SS-ELM) provides a way to fully utilize the easily acquired unlabeled data, the paper proposes a semi-supervised online sequential ELM, denoted as SOS-ELM. The proposed SOS-ELM not only has the advantage of learning in a sequential way, but also makes the most use of unlabeled data. Experiments have been done on benchmark problems of regression and classification and the results show that the proposed SOS-ELM outperforms OS-ELM in generalization performance with similar training speed and outperforms SS-ELM with much lower training time cost.