Tracing Geographical Origins of Teas Based on FT-NIR Spectroscopy: Introduction of Model Updating and Imbalanced Data Handling Approaches

Table 3

Improving the prediction of 2015 tea samples using three outlier detection based undersampling approaches.

Classifier	One Class SVM (OC-SVM)			Isolation forest (IF)			Elliptic envelope (EE)
	Recall 1	Recall 2	MAR^a	Recall 1	Recall 2	MAR	Recall 1	Recall 2	MAR
LDA	0.4993	0.3700	0.4346	0.4643	0.4100	0.4372	0.4498	0.3900	0.4199
SVM	0.7409	0.7300	0.7355	0.7191	0.7600	0.7395	0.7118	0.7500	0.7309
SGD^b	0.5997	0.5100	0.5549	0.5997	0.5100	0.5549	0.5662	0.5100	0.5381
Decision tree	0.6652	0.5700	0.6176	0.6958	0.5800	0.6379	0.6201	0.5400	0.5800
Random forest	0.7089	0.6300	0.6694	0.6812	0.6200	0.6506	0.6710	0.6100	0.6405
AdaBoost	0.6405	0.6800	0.6602	0.7322	0.5400	0.6361	0.6667	0.5600	0.6133
MLP^c	0.6856	0.5200	0.6028	0.8020	0.3800	0.5910	0.7205	0.3800	0.5503

^aMAR: macro average recall; ^bSGD: stochastic gradient descent; ^cMLP: multilayer perceptron.