Thank you for the answer. It really helps!

On Sat, Dec 15, 2012 at 3:15 AM, Ted Dunning <[email protected]> wrote:

> The point is that the AdaptiveLogisticRegression computes average
> performance over some number of training examples and then uses that
> average performance to adapt the algorithm hyper-parameters to get the best
> average results on held-out data.  You want a short window so that the
> average responds quickly, but you want a long window so that the average is
> less noisy.
>
> On Fri, Dec 14, 2012 at 11:03 AM, Yang Zhou <[email protected]> wrote:
>
> > The parameter of setInterval(int) is interval. And windowSize is the
> > parameter of setAveragingWindow(int). Sorry about the mistake.
> >
> > On Sat, Dec 15, 2012 at 2:52 AM, Yang Zhou <[email protected]> wrote:
> >
> > > I read the source code and know that windowSize for setInterval(int
> > > windowSize) is the number of training examples to use in optimization.
> > But
> > > I can still not understand what's the exact meaning of  windowSize for
> > > setAveragingWindow(int windowSize) . Would you mind telling more about
> > > that? Thanks!
> > >
> > >
> > > On Sat, Dec 15, 2012 at 2:44 AM, Ted Dunning <[email protected]
> > >wrote:
> > >
> > >> I would recommend testing with OnlineLogisticRegression first.
> > >>
> > >> The AdaptiveLogisticRegression has a tendency to freeze on sub-optimal
> > >> parameter values sooner than it should.
> > >>
> > >> In any case, the averaging window for ALR should be set fairly long
> and
> > >> should be at least 10% of your data set.  If your dataset is small, I
> > >> would
> > >> recommend using OLR instead.
> > >>
> > >>
> > >> On Fri, Dec 14, 2012 at 10:41 AM, Yang Zhou <[email protected]>
> > wrote:
> > >>
> > >> > Hi,
> > >> >
> > >> > I try to train a AdaptiveLogisticRegression, but have no idea how
> > large
> > >> > windowSize should be when calling setInterval(int windowSize) and
> > >> > setAveragingWindow(int windowSize) of AdaptiveLogisticRegression.
> Any
> > >> > suggestion? Thanks!
> > >> >
> > >>
> > >
> > >
> >
>

Reply via email to