Yes. 0 is perfect prediction. It can only be achieved by a score of 1 for the correct answer every time.
Note that average log-likelihood only works for probability scores. On Tue, May 31, 2011 at 6:38 PM, Xiaobo Gu <[email protected]> wrote: > On Tue, May 31, 2011 at 11:54 PM, Ted Dunning <[email protected]> > wrote: > > Argh.... > > > > log-likelihood should approach the percentage of INcorrect answers > > (negated). > > Then we just only to see if the average log likeliyhood is closer to 0 > to determine the perfmonce of the model, regardless the relationship > between it and percentage of INcorrect or correct answers? > > > > On Tue, May 31, 2011 at 7:49 AM, Xiaobo Gu <[email protected]> > wrote: > > > >> Page 228 of version 7 of Mahout in Action says : > >> > >> Log-likelihood has a maximum value of zero and no bound on how far > >> negative it can go. For highly accurate classifiers, the value of > >> average log-likelihood should be close to the average percent correct > >> for the classifier times the number of target categories. > >> > >> Average percent correct times the number of target categories is more > >> than 0, while Log-likelihood is always less than 0, then is the above > >> statement correct ? > >> > >> > >> > >> On Tue, May 31, 2011 at 10:45 PM, Xiaobo Gu <[email protected]> > >> wrote: > >> > Does it mean the percent of records that the model has correctlly > >> > predicted the target on the validate protion of the data set, then it > >> > should be between 0 and 1, and the bigger the better performance of > >> > the model ? > >> > > >> > Regards, > >> > > >> > Xiaobo Gu > >> > > >> > > >
