When to use LTR

[Ryan Yacyshyn <ry...@gmail.com>]

Hey all,

There's a lot of interest and articles explaining how learning to rank can
be implemented in search engines and that's awesome. I'm looking forward to
giving this a try.

However, I was wondering.. when would you know that implementing LTR
would *really
help* in any given case? Is it better for large indices, frequently updated
content, or more structured content, etc.? Or is it all trial and error?

Would be great to hear your thoughts on this.

Regards,
Ryan

Re: When to use LTR

[Ryan Yacyshyn <ry...@gmail.com>]

Thanks for your help Alessandro!

Ryan


On Wed, 21 Jun 2017 at 19:25 alessandro.benedetti <a....@sease.io>
wrote:

> Hi Ryan,
> first thing to know is that Learning To Rank is about relevancy and
> specifically it is about to improve your relevancy function.
> Deciding if to use or not LTR has nothing to do with your index size or
> update frequency ( although LTR brings some performance consideration you
> will need to evaluate) .
>
> Functionally, the moment you realize you want LTR is when you start tuning
> your relevancy.
> Normally the first approach is the manual one, you identify a set of
> features, interesting for your use case and you tune a boosting function to
> improve your search experience.
>
> e.g.
> you decide to weight more the title field than the content and then
> boosting
> recent documents.
>
> What happens next is :
> "How much should I weight more the title ?"
> "How much should I boost recent documents ?"
>
> Normally you just check some golden queries and you try to manually
> optimise
> these boosting factors by hand.
>
> LTR answers to this requirements.
> To make it simple LTR will bring you a model that will tell you the best
> weighting factors given your domain ( and past experience) to get the most
> relevant results for all the queries ( this is the ideal, of course it is
> quite complicated and it depends of a lot of factors)
>
> Of course it doesn't work like magic and you will need to extensively
> design
> your features ( features engineering), build a valid training set (
> explicit
> or implicit), decide the model that best suites your needs ( linear model
> or
> Tree based ?) and a lot of corollary configurations.
>
> hope this helps!
>
>
>
>
>
> -----
> ---------------
> Alessandro Benedetti
> Search Consultant, R&D Software Engineer, Director
> Sease Ltd. - www.sease.io
> --
> View this message in context:
> http://lucene.472066.n3.nabble.com/When-to-use-LTR-tp4342130p4342140.html
> Sent from the Solr - User mailing list archive at Nabble.com.
>

Re: When to use LTR

["alessandro.benedetti" <a....@sease.io>]

Hi Ryan,
first thing to know is that Learning To Rank is about relevancy and
specifically it is about to improve your relevancy function.
Deciding if to use or not LTR has nothing to do with your index size or
update frequency ( although LTR brings some performance consideration you
will need to evaluate) .

Functionally, the moment you realize you want LTR is when you start tuning
your relevancy.
Normally the first approach is the manual one, you identify a set of
features, interesting for your use case and you tune a boosting function to
improve your search experience.

e.g.
you decide to weight more the title field than the content and then boosting
recent documents.

What happens next is : 
"How much should I weight more the title ?"
"How much should I boost recent documents ?"

Normally you just check some golden queries and you try to manually optimise
these boosting factors by hand.

LTR answers to this requirements.
To make it simple LTR will bring you a model that will tell you the best
weighting factors given your domain ( and past experience) to get the most
relevant results for all the queries ( this is the ideal, of course it is
quite complicated and it depends of a lot of factors)

Of course it doesn't work like magic and you will need to extensively design
your features ( features engineering), build a valid training set ( explicit
or implicit), decide the model that best suites your needs ( linear model or
Tree based ?) and a lot of corollary configurations.

hope this helps!





-----
---------------
Alessandro Benedetti
Search Consultant, R&D Software Engineer, Director
Sease Ltd. - www.sease.io
--
View this message in context: http://lucene.472066.n3.nabble.com/When-to-use-LTR-tp4342130p4342140.html
Sent from the Solr - User mailing list archive at Nabble.com.