I'm working on a classification problem related to the marking of ip/tcp packet, the classes are Best Effort and Non Best Effort; I'm using Python language. I have selected these features: Protocol, Length of the packet, Port number used for source and destination, Ip addressess from source and destination, flag not fragment and the ECN field, everything to know a possible classification for the DSCP field.
My idea is to apply a dimensionality reduction for reduce the space where I'm working and see if it can improve my results (using Algorithms like Random Forest, Naive Bayes, SVM).
For now I have used only PCA that creates new axes and from them I can see the percentage of each variables considered respect to the beginning point.
However I have seen something related to the LDA and it maximes the variance according to the label, so it is supervidsed learning while PCA is unsupervised.
Finally what do you suggest to me, how can I proceed ? Cause I do not know when I have to use PCA for improving classification results and when LDA . And if is correct to use them.
