I am working on the dataset KDD'99. It has 4 main classes, 39 features and about 3 million instances with very unbalance distribution over classes. As a newbie, I am curious about what are correct steps to take for classification on these 4 classes.
I am pointing my planned steps, you may also indicate and mistake on my flow.
- Take a subset of the dataset (but don't how how small)
- Feature elimination by considering unvarying features with some variance threshold
- Mean cancellation and Covariance equalization.(Whitening transform can be used for reduction of correlation of features) (May be reduction of features if some of them have over correlation.)(You may suggest PCA but I need to infer the results at the ends to see the importance of the features)
- I aim to shuffle instances to have most informative sequence to classification algorithm (since the data set is big the time of convergence maybe decreased) but I don't know any algorithm.
- Then using any fast converged algorithm (stochastic ones maybe) classify.