log2g at [−10, 10]. Results showed

log2g at [−10, 10]. Results showed that the best crossvalidation
accuracy was obtained with log2c=−10 and
log2g=−10, corresponding to c=0.001 and g=0.001. The
result (Table 2) showed that the accuracies of the calibration
set from electronic tongue, nose, and the combination of both
were 96, 98, and 99 %, respectively, and those of the prediction
set were 70, 75, and 85 %, respectively.
Classification Using KNN
KNN model was trained by data from 100 samples for the
training set, calibrated by comparing to the true type and used
to predict 20 other samples for the testing set. This model was
corresponded to the data from electronic tongue, nose, and the
combination of both, which were reduced by PCA and LLE
algorithms, respectively.
In PCA-KNN model, Fig. 5b shows the error value of
KNN model by cross-validation according to different K
values in the calibration set. As can be seen from this figure,
the best cross-validation accuracy were achieved when K=3.
The result showed that the accuracies of the calibration set
from electronic tongue, nose, and the combination of both
were 94, 97, and 95 %, respectively, and those of the prediction
set were 45, 75, and 55 %, respectively (Table 2).
In LLE-KNN model, as can be seen from Fig. 5b, the error
value of KNN model by cross-validation according to different
K values in the calibration set displayed no difference. In
this paper, K=3 was chosen. The result (Table 2) showed that
the accuracies of the calibration set from electronic tongue,
A
0.200
0.250
0.300
0.350
0.400
0.450
5 6 7 8 9 10 11 12 13 14 15
k values
Error values %