For example, i calculate the mean and variance of an EEG signal, then the mean and variance are two features.
I wish to evaluate the performance of each of these features for distinguishing between normal and seizure.
Now, would the ROC curve be different for the feature,say,average, for a linear classifier(LDA) from the ROC obtained for average for a SVM classifier?
It should be different. What ROC curve represents is the trade-off between sensitivity and specificity means how accurate the classification is. So if you apply two different classifiers on the same feature space then they will yield different accuracy measures and hence different ROC curves.
Yes it might be possible that u get different ROC curve from different classifier... but the variation of curves would be not so different.. Its better to use the same classifier that has been used in previous related works...
There might be difference and thats why ull need to choose at last more appropriate classifier that exposes the structure in ur data best to learning algorithm
The ROC curve for different feature-set would also be different and for different classifiers also it will be different achieving different GAR and FAR. Therefore, select an optimal combination of feature-set and classifier which gives best accuracy to get ROC curve.
The only thing you need for building a ROC curve is a score. Therefore, the score can be either the decision function of the classifier, or directly a feature. Then indeed, with two different classifier, the decision function will change together with the associated ROC curve.
To determine ROC you need a rigid scoring system for the applied cognitive and classification system. The produced ROC depends the type of used classifier.
But for optimal classifiers the shape of produced ROCs will be similar up to certain level. Regards
for feature evaluation you can use the inter & intra scatter analysis to asses its discrimination capability... Also the you can use the co-variance matrix to avoid the selecting correlated features even though each have good discrimination power, this will avoid the occurrence of decision biasing.
Yes, you can analyze the performance of learning/testing using ROC(receiver operating characteristics).
But it is not only the way and not mandatory.
This might be varied from classifier to classifier on the same data samples.
ROC curve will be different for different classifier. But when you are using the different classifier and score obtained are similar then their should be some similarity between the curve as suggested by Prof. Loay E. George.
Dear Shanir,
It is a nice question indeed. The answer is that results would vary be different. The AUC value you can take for the selection of the best classifier.
Mr, Shanir,
ROC Curve tell about the perforance of the system,
also dont lie only on mean and variance to detect the signal as normal and abnormal.. collect more feature that will be use full to discriminate.
- Badges
- Science topic
- Similar topics
- Filing