Adaptive Online Sequential ELM for Concept Drift Tackling

Table 8

Average testing accuracy in % for RD experiment. For SEA, we started from to . For MNIST, we started from to .

(a) Sudden drift effect caused by split training composition with hidden nodes increase

Data	AOS-ELM		Concept	Testing accuracy
SEA		0		90.00 ± 0.59
				90.24 ± 0.6
		500		66.29 ± 1.12
				90.12 ± 0.52
MNIST		0		96.42 ± 0.21
				93.68 ± 0.23
		500		17.59 ± 0.98
				97.08 ± 0.15

(b) MNIST RD simulation: the effect of hidden nodes increase for split and shuffled training composition ()

Data		Composition
MNIST	0	Split	96.42 ± 0.21	93.68 ± 0.23
		Shuffled	96.09 ± 0.12	94.34 ± 0.16
	500	Split	17.59 ± 0.98	97.08 ± 0.15
		Shuffled	96.53 ± 0.12	94.29 ± 0.25
	1000	Split	8.65 ± 1.13	97.64 ± 0.18
		Shuffled	96.74 ± 0.14	94.78 ± 0.10