why better?

[jovo]

if i recall, RerF is 10% better than RF on about 10% of the data?

let's compute, for each dataset:

1. n
2. p
3. p/n
4. sum of singular values
5. sum of squared singular values

let's make a pairs-plot, 5 x 5 panels, color code by much better than RF (eg >7% or so), and not. and see if we can see anything?

in the PAMI paper, we really should try to answer the questions:

1. which features/subspaces were informative
2. why does RerF > RF (in terms of bias and variance)
3. what properties of data do we expect RerF > RF