Online Sequential Projection Vector Machine with Adaptive Data Mean Update

Table 2

Comparison of OSPVM, Batch-PVM, BSGD, AMM, and Pegasos.

Dataset	Algorithms	Nodes ()	Training time (s)	Testing time (s)	Training accuracy	Testing accuracy
Face	OSPVM (40, 16-by-16)	51 (0.96)	1.50 s	0.0004 s	99.89%	92.87%
	OSPVM (40, 1-by-1)	43 (0.99)	13.07 s	0.0005 s	99.20%	91.20%
	Batch-PVM	65	0.460 s	0.0005 s	99.81%	92.30%
	SVD + BSGD [10]	200	1.542 s	0.0835 s	99.92%	91.63%
	SVD + AMM Online [13]	200	1.990 s	0.0300 s	99.82%	88.75%
	SVD + Pegasos [12]	—	1.530 s	0.0240 s	99.11%	86.38%
Secom	OSPVM (40, 16-by-16)	61 (0.96)	1.67 s	0.007 s	94.08%	93.14%
	OSPVM (40, 1-by-1)	16 (0.96)	4.01 s	0.0004 s	94.14%	93.3%
	Batch-PVM	60	0.525 s	0.0073 s	93.37%	93.35%
	SVD + BSGD	100	1.801 s	0.0083 s	95.12%	93.13%
	SVD + AMM Online	100	12.19 s	0.031 s	94.11%	87.87%
	SVD + Pegasos	—	1.660 s	0.026 s	93.16%	89.12%
Arcene	OSPVM (40, 16-by-16)	106 (0.96)	61.17 s	0.0005 s	95.88%	90.50%
	OSPVM (40, 1-by-1)	39 (0.96)	130.6 s	0.0004 s	93.5%	86.7%
	Batch-PVM	85	5.06 s	0.00038 s	94.63%	90.80%
	SVD + BSGD	200	65.22 s	0.0335 s	95.92%	90.43%
	SVD + AMM Online	200	81.69 s	0.06 s	94.89%	87.75%
	SVD + Pegasos	—	56.41 s	0.044 s	94.42%	86.31%
Dexter	OSPVM (40, 16-by-16)	176 (0.96)	131.1 s	0.004 s	97.88%	92.25%
	OSPVM (40, 1-by-1)	86 (0.96)	619.3 s	0.004 s	96.0%	91.20%
	Batch-PVM	160	10.36 s	0.005 s	98.38%	91.25%
	SVD + BSGD	200	148.54 s	0.003 s	97.98%	92.63%
	SVD + AMM Online	200	178.19 s	0.003 s	96.81%	89.95%
	SVD + Pegasos	—	119.40 s	0.004 s	95.87%	87.36%
Multi.fea.	OSPVM (40, 16-by-16)	55 (0.96)	4.93 s	0.0053 s	98.16%	94.40%
	OSPVM (40, 1-by-1)	38 (0.96)	13.4 s	0.0047 s	96.6%	93.4%
	Batch-PVM	160	1.83 s	0.0192 s	99.98%	95.67%
	SVD + BSGD	200	5.54 s	0.0095 s	98.42%	94.63%
	SVD + AMM Online	200	10.79 s	0.03 s	99.82%	92.15%
	SVD + Pegasos	—	4.46 s	0.034 s	99.82%	91.88%
News20	OSPVM (40, 16-by-16)	1110 (0.96)	1283 s	19.8 s	85.26%	83.10%
	OSPVM (40, 1-by-1)	1100 (0.96)	1949 s	19.9 s	85.6%	83.14%
	Batch-PVM	1000	1060 s	19.2 s	84.89%	83.12%
	SVD + BSGD	1200	2289 s	18.6 s	83.52%	82.33%
	SVD + AMM Online	1200	2679 s	21.3 s	83.83%	82.25%
	SVD + Pegasos	—	1679 s	19.2 s	83.22%	81.81%
Sector	OSPVM (40, 16-by-16)	130 (0.96)	10.12 s	0.20 s	88.86%	78.40%
	OSPVM (40, 1-by-1)	150 (0.96)	18.4 s	0.21 s	86.6%	79.04%
	Batch-PVM	160	2.13 s	0.21 s	87.98%	79.01%
	SVD + BSGD	200	7.53 s	0.34 s	87.44%	76.68%
	SVD + AMM Online	200	12.69 s	0.33 s	86.81%	76.65%
	SVD + Pegasos	—	6.45 s	0.34 s	86.12%	75.88%

Note: since OSPVM is equivalent to PVM rather than an approximation,  if it has the same experimental setting (same number of hidden nodes and same training and testing splits), OSPVM and PVM would obtain the same performance (training accuracy and testing accuracy).