I have KDD dataset for detecting fraud actions on networks but it has millions of lines and >20 feature columns. Thus it is not viable to process all these on my personal computer. I am thinking about the random sampling on data to reduce the rows to some legible numbers than I want also to reduce the number of features. I know PCA can be used to reduce the feature number but first I also want to get rid of the some dirty features that are not helpful enough to classification purpose.
What are the algorithms can be used for that feature elimination and extraction purpose?