Argh... make that last be w(a-b) log w(a-b)/\gamma
On Sun, Feb 23, 2014 at 3:42 PM, Ted Dunning <ted.dunn...@gmail.com> wrote: > > On Sun, Feb 23, 2014 at 10:51 AM, Frank Scholten (JIRA) > <j...@apache.org>wrote: > >> Ted: can you tell a bit more about the log transforms? Some of them are >> just Math.log while others are more complex expressions. > > > The increased complexity comes up when there are zero or small negative > values. > > In general, monetary values are commonly transformed with a log during > training of a logistic regression model. Often you retain the original as > well. > > The motivation for the log is that it is common for the structure of the > problem to depend as much on relative differences rather than absolute > differences. Thus, $80 is different from $100 in about the same way that > $800 is different from $1000. This makes sense if you are talking about > what makes a material difference. > > Of course, if you are talking about net profits, then you may want > features that look like log(a-b) instead. What happens when that goes > negative is a bit of a can of worms in terms of feature design. Sometimes, > a small reference value is defined and a value like w(a-b) log w(a-b) is > used where w(x) = x-\gamma if x > \gamma, x+\gamma if x < -\gamma and 0 > else. > > >
