Thanks for the added input.
I'll certainly look into the machine learning aspect, will be good to
put some basic knowledge I have into practice.
I'd been led to believe the tie parameter didn't actually do a lot. :-/
On 03/18/2016 12:07 PM, Nick Vasilyev wrote:
I work with a similar catalog; except our data is especially bad. We've
found that several things helped:
- Item level grouping (group same item sold by multiple vendors). Rank
items with more vendors a bit higher.
- Include a boost function for other attributes, such as an original image
of the product
- Rank items a bit higher if they have data from an external catalog like
IceCat
- For relevance and performance, we have several fields that we copy data
into. High value fields get copied into a high weighted field, while lower
value fields like description get copied into a lower weighted field. These
fields are the backbone of our qf parameter, with other fields adding
additional boost.
- Play around with the tie parameter for edismax, we found that it makes
quite a big difference.
Hope this helps.
On Fri, Mar 18, 2016 at 6:19 AM, Alessandro Benedetti <abenede...@apache.org
wrote:
In a relevancy problem I would repeat what my colleagues already pointed
out :
Data is key. We need to understand first of all our data before we can
understand what is relevant and what is not.
Once we specify a groundfloor which make sense ( and your basic approach +
proper schema configuration as suggested + properly configured request
handler , seems a good start to me ) .
At this point if you are still not happy with the relevancy (i.e. you are
not happy with the different boosts you assigned ) my strongest suggestion
at this time is to move to machine learning.
You need a good amount of data to feed the learner and make it your Super
Business Expert) .
I have been recently working with the Learn To Rank Bloomberg Plugin [1] .
In my opinion will be key for all the business that have many features in
the game, that can help to evaluate a proper ranking.
For that you need to be able to collect and process signals, and you need
to carefully tune the features of your interest.
But the results could be surprising .
[1] https://issues.apache.org/jira/browse/SOLR-8542
[2] Learning to Rank in Solr <https://www.youtube.com/watch?v=M7BKwJoh96s>
Cheers
On Thu, Mar 17, 2016 at 10:15 AM, Robert Brown <r...@intelcompute.com>
wrote:
Thanks Scott and John,
As luck would have it I've got a PhD graduate coming for an interview
today, who just happened to do her research thesis on information
retrieval
with quantum theory and machine learning :)
John, it sounds like you're describing my system! Shopping products from
multiple sources. (De-duplication is going to be fun soon).
I already copy fields like merchant, brand, category, to string fields to
use them as facets/filters. I was contemplating removing the description
due to the spammy issue you mentioned, I didn't know about the
RemoveDuplicatesTokenFilterFactory, so I'm sure that's going to be a huge
help.
Thanks a lot,
Rob
On 03/17/2016 10:01 AM, John Smith wrote:
Hi,
For once I might be of some help: I've had a similar configuration
(large set of products from various sources). It's very difficult to
find the right balance between all parameters and requires a lot of
tweaking, most often in the dark unfortunately.
What I've found is that omitNorms=true is a real breakthrough: without
it results tend to favor small texts, which is not what's wanted for
product names. I also added a RemoveDuplicatesTokenFilterFactory for the
name as it's a common practice for spammers to repeat some key words in
order to be better placed in results. Stemming and custom stop words
(e.g. "cheap", "sale", ...) are other potential ideas.
I've also ended up in removing the description field as it's often too
broad, and name is now the only field left: brand, category and merchant
(as well as other fields) are offered as additional filters using
facets. Note that you'd have to re-index them as plain strings.
It's more difficult to achieve but popularity boost can also be useful:
you can measure it by sales or by number of clicks. I use a combination
of both, and store those values using partial updates.
Hope it helps,
John
On 17/03/16 09:36, Robert Brown wrote:
Hi,
I currently have an index of ~50m docs representing shopping products:
name, description, brand, category, etc.
Our "qf" is currently setup as:
name^5
brand^2
category^3
merchant^2
description^1
mm: 100%
ps: 5
I'm getting complaints from the business concerning relevancy, and was
hoping to get some constructive ideas/thoughts on whether these boosts
look semi-sensible or not, I think they were put in place pretty much
at random.
I know it's going to be a case of rounds upon rounds of testing, but
maybe there's a good starting point that will save me some time?
My initial thoughts right now are to actually just search on the name
field, and maybe the brand (for things like "Apple Ipod").
Has anyone got a similar setup that could share some direction?
Many Thanks,
Rob
--
--------------------------
Benedetti Alessandro
Visiting card : http://about.me/alessandro_benedetti
"Tyger, tyger burning bright
In the forests of the night,
What immortal hand or eye
Could frame thy fearful symmetry?"
William Blake - Songs of Experience -1794 England
