Marwa A.

If you applied PCA correctly in the "Classify" tab (presuming that you are using WEKA's Explorer), then you will find the selected attributes in the "Classifier output" panel in the "Classify" tab, where there will be a section called "Selected attributes".

HTH.

Dr. Samer Sarsam

I did what you mentioned but I got vector of features in the first image attached. the other algorithm I'm comparing with gave me features result in the second attached image. I have the accuracy and precision higher that him but how I can tell which features subset is the best. How to deal with the vectors I got in order to give conclusions

Marwa A.

Thanks for providing the figures which showed that you are not doing the right process. First, you are not applying PCA; you are using “InfoGainAttributeEval” and “CfsSubsetEval” methods. Second, you are using the “Select attributes” tab, which is the wrong tab, in your case.

If you need to apply PCA together with classification process to find the selected features. You need to use “PrincipalComponents” (attribute evaluator) + Ranker (search method) together from the “Classify” tab instead of the “Select attributes” tab. To do so, from the “Classify” tab, choose “AttributeSelectedClassifier” and specify its 3 parameters: classifier, evaluator, and search as J48 (you can select your desired classifier), PrincipalComponents, and Ranker, respectively. Make sure that you do this process with the use of 10-fold cross-validation. Now, you will find the selected features (here the features are ranked by the Ranker that ranks attributes by the score produced for each by the evaluator) in the “Selected attributes” section in the “Classifier output” tab. You can apply this procedure several times and change only the classifier, then assess your result via accuracy and ROC, for instance. The best classifier produces the highest accuracy and ROC results.

Cheers,

Dr. Samer Sarsam

Hi Dr Samer ,

I think there could be a small misunderstanding. Actually I applied the PCA on my data(25 features) through these steps I got from stackoverflow.com

Classify> Meta->FilteredClassifier classifier. Set the filter to PrincipalComponents and and the 4 different classifiers I chose through PCA. The result in first figure shows the c4.5 classifier.

The Second figure is paper experiment I'm comparing my model with. It used ( cfs - c4.5 - NB - RF- Reptree ) in some way for feature selection. I'm attaching fig.2 just to show that the classifier's result of the paper I'm comparing my model with is simply subset of features from the original dataset.

My question is while using PCA I got vector of features which are considered as new features and has different type from the original features. If PCA shows good precision and accuracy in classification than the previous model I'm comparing with .. I need to specify which features are the most important for detecting an attack ( like in fig.2 ).

how to say that if PCA gave vector of features not specific subset of features.