Hello Ilay,
Answers in line:
On Sat, Nov 18, 2017 at 2:22 PM, ilay <ilay....@gmail.com> wrote:
>
> 1. Does LTR only support phrase matching (complete user query) from training
> data for extracting feature score:
> ex.
> efi.user_query='tv+stand' matches the title feature only if title contains
> "tv stand" in the title.
> By removing the quotes, able to match at term level, but the behaviour is
> not consistent when we change the order of the terms in the query.
> i.e. efi.user_query=tv stand gives a different feature match score that of
> efi.user_query=stand tv for the same title match.
>
> Are we supposed to always wrap efi.userquery with single quotes and do
> phrase matching. If we do so, we miss out on term matches.
> Which request handler does this query go through?
I'm not sure I understand properly the question, could you post the
feature that you are using together with the efi parameter? I guess
you have a feaure like this:
{
"name" : "myFeature",
"class" : "org.apache.solr.ltr.feature.SolrFeature",
"params" : {
{"q":"$user_query"}
}
},
with efi.user_query='tv+stand' you get the score of the boolean query
with the terms in and while efi_query='tv OR stand' should give you a
score of the OR (so it will give you a score different from zero if
only one of the two terms is in the field). If you want to match
exactly the bigram in the text I think efi.user_query='\"tv+stand\"'
should work. Let me know if it solves.
>
> 2. Generating training data using clickstream
>
> Please advice on usage of clickstream data for training (in the absence of
> human judgements).
> Can we expect LTR to do good job interms of weights learned when we use
> click data (implicit feedback data).
This really depends on your use case. I would suggest to take a look
at this survey by Chuklin et al.
clickmodels.weebly.com/uploads/5/2/2/5/52257029/mc2015-clickmodels.pdf
>
> 3. Newness challenge
> Generally clicks data is good for popular items. Learning newness seems a
> challenge with this approach. Any thoughts here..
Not really, i'm afraid it's an open problem, crowdsourcing may help..
> 4. Original score feature weight is still zero
Did you take a look at the original scores? there was a bug that was
making solr always return zero instead of the original score, so maybe
that's your problem.
The bug has been recently fixed in
https://issues.apache.org/jira/browse/SOLR-11180. It would explain
why your weight is zero.
Best,
Diego
>
> --
> Sent from: http://lucene.472066.n3.nabble.com/Solr-User-f472068.html
