madlib-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Srivatsan Ramanujam <vatsan...@utexas.edu>
Subject Elastic net regression - log likelihood
Date Tue, 08 Dec 2015 17:55:59 GMT
Hi guys,

I was looking at this:
http://doc.madlib.net/latest/group__grp__elasticnet.html
and noticed that the log_likelihood is defined as the *negative* value of
the L(w) defined in the equation.

Any reason why you've included the negative sign?

I was a little confused as i assumed the log_likelihood reported was
actually the value of L(w) (the lower the better since we are minimizing
L(w) for w), but then noticed while doing a grid-search on the alpha and
lambdas my best performing models had log_likelihood values < 0 but close
to zero. The values which had higher magnitude (negative sign again)
performed poorly. Just wondering what's the need to multiply the
log_likelihood by -1, why not report raw values and let user pick the w
with lowest L(w) ?

Please clarify if i have misunderstood.

Thanks
Vatsan

Mime
View raw message