Semi-Supervised Online Elastic Extreme Learning Machine with Forgetting Parameter to deal with concept drift in data streams
Carlos A. S. da Silva, Renato Antonio Krohling · 2019
Concept drift is a common problem for online sequential algorithms that deal with data streams. Many supervised and unsupervised approaches to solve concept drifts were proposed recently, including some ELM-based algorithms. Due to its fast training, ELM-based algorithms can quickly adapt to dataset changes, detecting and preventing concept drift. SSOE-ELM is a semi-supervised online ELM-based algorithm with good accuracy and generalization ability, but as an online algorithm it is also affected by the concept drift problem. In this paper, a variation of SSOE-ELM algorithm with a semi-supervised concept drift detector and a forgetting parameter called SSOE-FP-ELM is proposed. This new approach is compared with standard SSOE-ELM and FP-ELM. Our experimental results show that SSOE-FP-ELM outperforms SSOE-ELM and FP-ELM in accuracy with two different concept drift types, without a considerable increase in training time.