Spyros, thank you for verifying this, we are planning to do something
similar.
Thanks,
Roopa
On Fri, Jul 16, 2021 at 12:09 PM Spyros Kapnissis <ska...@gmail.com> wrote:
> Hello,
>
> Just to verify this, we had come across the exact same issue when
> converting an XGBoost model to MUltipleAdditiveTrees. This was an issue
> specifically with the categorical features that take on integer values. We
> ended up subtracting 0.5 from the threshold value on any such split point
> on the converted model, so that it would output the same score as the input
> model.
>
> On Fri, Jul 16, 2021, 18:19 Roopa Rao <roop...@gmail.com> wrote:
>
> > Okay, thank you for the input
> >
> > Roopa
> >
> > On Fri, Jul 16, 2021 at 5:55 AM Alessandro Benedetti <
> a.benede...@sease.io
> > >
> > wrote:
> >
> > > Hi Roopa,
> > > I was not able to find why that slack was added.
> > > I am not sure why we would like to change the threshold.
> > > I would recommend creating a Jira issue and tag at least myself,
> > Christine
> > > Poerschke and Diego Ceccarelli, so we can discuss and potentially open
> a
> > > pull request.
> > >
> > > Cheers
> > >
> > >
> > > --------------------------
> > > Alessandro Benedetti
> > > Apache Lucene/Solr Committer
> > > Director, R&D Software Engineer, Search Consultant
> > >
> > > www.sease.io
> > >
> > >
> > > On Thu, 15 Jul 2021 at 22:24, Roopa Rao <roop...@gmail.com> wrote:
> > >
> > > > Hi All,
> > > >
> > > > In LTR for MultipleAdditiveTreeModel what is the purpose of adding
> > > > NODE_SPLIT_SLACK
> > > > to the threshold?
> > > >
> > > > Reference: org.apache.solr.ltr.model.MultipleAdditiveTreesModel
> > > >
> > > > private static final float NODE_SPLIT_SLACK = 1E-6f;
> > > >
> > > >
> > > > public void setThreshold(float threshold) { this.threshold =
> threshold
> > +
> > > > NODE_SPLIT_SLACK; }
> > > >
> > > > We have a feature which can return 0.0 or 1.0
> > > >
> > > > And model with this tree:
> > > >
> > > >
> > is_xyz_feature,threshold=0.99999994,left=0.0010180053,right=-0.0057609854
> > > >
> > > > However when Solr actually scores it it is taking it as follows
> > > > is_xyz_feature:1.0<= 1.000001, Go Left
> > > >
> > > > So all the time it goes to left which is incorrect.
> > > >
> > > > Thanks,
> > > > Roopa
> > > >
> > >
> >
>
