3. FOOD ChEMISTRy & bIOTEChNOLOGy 3.1. Lectures
3. FOOD ChEMISTRy & bIOTEChNOLOGy 3.1. Lectures
3. FOOD ChEMISTRy & bIOTEChNOLOGy 3.1. Lectures
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Chem. Listy, 102, s265–s1311 (2008) Food Chemistry & Biotechnology<br />
case of using the leave-three-out validation. Thus, the remaining<br />
one sample or three samples were not included in the<br />
training procedure but were used for inspecting the quality<br />
of prediction whether the predicted variety of wine matches<br />
the real wine variety. The results achieved by several classification<br />
techniques and different software are summarized<br />
in Table I. The presented Ann results were achieved after<br />
optimising the neural network; the lowest error was obtained<br />
when using a three layer perceptron with 30 input neurons<br />
(areas of the selected best peaks), 3 hidden neurons and<br />
one six-level output neuron (representing the predicted wine<br />
variety).<br />
All wine samples (100 %) were correctly classified into<br />
six classes by variety when the calculated multidimensional<br />
model is considered – no one sample was allocated to a wrong<br />
class. Considering the validation results, the performance in<br />
leave-one-out validation depend on the applied multivariate<br />
technique. The leave-three-out manual technique can be used<br />
also in the case when the automatic leave-one-out validation<br />
is not enabled for the given method and software. All validation<br />
results shown in Table I are above 90 %, which justifies<br />
very good ways of wine variety prediction enabling to confirm<br />
or reject wine authenticity.<br />
C l a s s i f i c a t i o n o f D r i n k i n g W a t e r<br />
Three classification criteria were used for the classification<br />
of drinking waters: (1) by three types of water – potable,<br />
mineral and spring water, (2) by five types of water – potable,<br />
mineral, mineral carbonated, spring and spring carbonated<br />
water, (3) by three countries of origin – Slovenia, Croatia<br />
and Czechia (the category of French waters was not used due<br />
to a very low number of samples). Before the calculations<br />
new categorical variables were created, which correspond to<br />
the first, second and third classification criterion: WType 3,<br />
WType 5 and Country, resp. In addition, a special categorical<br />
variable Carbon was created in order to mark whether the<br />
sample is carbonated or not (c/n). When using the Anns it is<br />
possible to utilize this variable at the input to provide some<br />
Table I<br />
Success in prediction of wine variety using 72 varietal wines,<br />
30 optimally selected chromatographic peaks, 5 classification<br />
and 2 validation techniques<br />
Method Classif. Leave-1- Leave-3- Software<br />
success out out<br />
LDA<br />
Correct/all<br />
%<br />
69/72<br />
95.8<br />
69/72<br />
95.8<br />
QDA<br />
correct/all<br />
%<br />
66/72<br />
91.7<br />
65/72<br />
90.3<br />
Knn<br />
correct/all<br />
%<br />
72/72<br />
100.0<br />
67/72<br />
9<strong>3.</strong>1<br />
LR<br />
correct/all<br />
%<br />
–<br />
–<br />
69/72<br />
95.8<br />
Ann<br />
correct/all<br />
%<br />
–<br />
–<br />
68/72<br />
94.4<br />
SAS<br />
SAS<br />
SAS<br />
SPSS<br />
JMP<br />
s558<br />
additional information about the sample. It is worth noting<br />
that the discriminant analysis techniques, except logistic<br />
regression, do not permit the use of non continuous input<br />
variables. The mentioned additional information cannot be<br />
of course used when the classification by the second criterion<br />
is used.<br />
Table II shows the classification results for five cases<br />
using categorization by water type into 3 and 5 classes, the<br />
same categorization but with the help of additional categorical<br />
variable Carbon, and finally categorization by 3 countries<br />
of origin. Intelligent Problem Solver is an extremely useful<br />
module of Trajan software facilitating the selection of the<br />
optimal neural network. For the sake of place, Table II exhibits<br />
only five best networks, automatically selected by this<br />
module, but a good possibility is to make a choice among<br />
a larger nuber of networks. Moreover, the networks belonging<br />
to different Ann variants can be examined in this way<br />
(e.g. Radial Base Ann). In Table II, the ordinal number of<br />
the network is marked by no, the number of neurons in individual<br />
layers is marked by I (input), H (hidden), and O (output);<br />
Err. indicates the sum of squares error obtained both<br />
for the training and test sets. The most important results are<br />
Table II<br />
Ann – selection of the best network for 5 different criteria of<br />
water classification using Intelligent Problem Solver of software<br />
Trajan 6.0<br />
MLP networks<br />
Categor<br />
neurons<br />
variables<br />
no I H O<br />
Train.<br />
set<br />
Err.<br />
Test<br />
set<br />
Err.<br />
Success<br />
[%]<br />
1 31 6 1(3) 0.009 <strong>3.</strong>759 77.3<br />
2 31 6 1(3) 0.023 4.706 81.8<br />
Country 3 31 5 1(3) 0.023 2.801 77.3<br />
4 31 6 1(3) 0.017 <strong>3.</strong>464 81.8<br />
5 31 6 1(3) 0.0044 5.239 86.4<br />
1 32 5 1(3) 0.0079 <strong>3.</strong>215 86.4<br />
Country<br />
c/n<br />
2 32<br />
3 32<br />
4 32<br />
6<br />
6<br />
6<br />
1(3)<br />
1(3)<br />
1(3)<br />
0.001<br />
0.100<br />
0.0032<br />
<strong>3.</strong>921<br />
2.242<br />
1.829<br />
81.8<br />
86.4<br />
86.4<br />
5 32 6 1(3) 0.144 0.480 90.9<br />
6 31 6 1(3) 0.0054 0.0026 100.0<br />
7 31 6 1(3) 0.0008 1.602 86.4<br />
WType3 8 31 6 1(3) 0.0002 0.649 95.5<br />
9 31 6 1(3) 0.0010 0.204 95.5<br />
10 31 6 1(3) 0.0000 0.935 86.4<br />
1 32 6 1(3) 1.4 × 10<br />
WType3<br />
–5 2. × 10 –5 100.0<br />
c/n<br />
2 32 6 1(3) 1.1 × 10 –6 5.8 × 10 –6 3 32 5 1(3)<br />
100.0<br />
5.9 × 10 –6 4 32 6 1(3)<br />
0.0029 100.0<br />
1.2 × 10 –9 8.2 × 10 –6 100.0<br />
5 32 6 1(3) 1.5 × 10 –8 9.1 × 10 –3 100.0<br />
1 31 6 1(5) 0.2946 0.744 90.9<br />
2 31 6 1(5) 0.215 0.599 86.4<br />
WType5 3 31 6 1(5) 0.424 1.154 81.8<br />
4 31 6 1(5) 0.197 0.4498 59.1<br />
5 31 6 1(5) 0.480 0.934 90.9