25.12.2013 Views

CRANFIELD UNIVERSITY Eleni Anthippi Chatzimichali ...

CRANFIELD UNIVERSITY Eleni Anthippi Chatzimichali ...

CRANFIELD UNIVERSITY Eleni Anthippi Chatzimichali ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

2.3.2.2 Class Prediction Accuracies<br />

Even though the overall accuracy ( ) of a classification model is of the utmost<br />

importance, the success of a classifier may be also assessed by a plethora of other<br />

performance metrics. In a multi-class case, it is interesting to investigate class<br />

predictions, which determine how well a sample belonging to a specific class has<br />

been predicted. The percentages of correctly classified samples per class and<br />

instrument are summarised in Figure 2-7. The comparison of the class predictions<br />

presented in the figure confirms that the overall accuracies of a classifier may be<br />

occasionally misleading. For instance, even though the overall accuracies of a linear<br />

and nonlinear classifier may appear to be similar, a closer inspection of the class<br />

predictions may reveal significant differences among the different classifiers.<br />

As presented in Section 2.2.2, the data intersection approach extracted for case<br />

study 1 a total of 32 common samples along with their respective sensory scores;<br />

these samples consist of 10 fresh (F), 6 semi-fresh (SF) and 16 spoiled (S) samples. In<br />

this instance, the spoiled samples constitute the majority class, whereas semi-fresh<br />

samples the minority class.<br />

Even though HPLC generated the highest overall accuracies ( ) among the<br />

datasets of case study 1, FTIR presented the strongest and equally stable class<br />

predictions among all classification models, with noteworthy better rates for semifresh<br />

samples than the remaining techniques. The FTIR data presented better overall<br />

as well as per-class prediction rates for the linear PLS-DA classifiers via<br />

bootstrapping or LOOCV; this is verified by both the overall accuracies of Figure 2-6<br />

and the class predictions of Figure 2-7. As thoroughly discussed in the previous<br />

section, it may be the case that the FTIR samples are linearly separable, thus, the<br />

complex boundaries of an SVM may not be able to separate them as accurately.<br />

Furthermore, it is notable that the classifiers, and especially the nonlinear SVMs, are<br />

relatively biased towards the majority class resulting in higher percentages of<br />

correctly classified spoiled samples.<br />

50

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!