On the scalability of feature selection methods on high-dimensional data