For importance it's probably best to stick with absolute values of
coefficients, instead of value of the penalty parameter for which the
coefficients changed to non-zero.
Friedman skipped a lot of details on his rule ensemble in that talk, due to
time constraint. In his implementation he was
Attended JSM last week and Friedman mentioned the use of LASSO for
variable selection (he uses it for rules ensembles). I am an
econometrician and not familiar with, i started running the examples in
R this week and you get to the plots section of the LARS package.
Plots of beta/max(beta)
Zubin,
my understanding about lasso is that it is a restricted version of
regression, where minimization of sse subject to sum(abs(beta)) upper
limit such that for unimportant feature, its beta will be restricted by
ZERO. the whole game of lasso is to find the proper upper limit. I think in
