Erik Bernhardson <ebernhardson@wikimedia.org> schrieb am Mi., 14. März 2018 um 15:34 Uhr:
Sorry for the delayed response, I've been out the last week. Responses inline.On Mon, Mar 12, 2018 at 1:27 AM, Georg Sorst <g.sorst@findologic.com> wrote:Erik,is there some documentation / further reading available on the machine ranking used for Wikipedia? This sounds very interesting!The code for managing all the data and training models is in https://github.com/wikimedia/search-mjolnir. This is a pyspark application that starts with the logged click data and transforms it into trained models. The models are currently trained using xgboost, but we are considering lightgbm as a replacement. Collecting click data is done separately with some processing of web request logs to match up search requests with their clicks.
Great stuff, thank you!
And can you elaborate on how the aggregated search queries are PII?The problem is that any aggregation of search queries that wants to be used to learn a ranking function needs to be provided the original query string. That string is then not aggregated, it is passed straight through from the users keyboard to the output data. We unfortunately don't have the kind of search volume, and don't keep long enough records (only 90 days) , to place arbitrary limits for minimum unique sessions issuing a query,and still have data that is representative of the whole. For example on english wikipedia, which is by far the most popular, only 60% of search sessions involve a query that was issued more than 10 times in the last 90 days. And 10 times is *way* too low for public release (I'm not sure where a reasonable cutoff might be, but its certainly not 10).
Thank you!GeorgGeorg Sorst <g.sorst@findologic.com> schrieb am Mo., 5. März 2018 um 20:31 Uhr:Awesome! Is there more info available somewhere? Algorithms used etc. maybe even source code?Erik:Is the raw data available? Can I compute the clickstream myself?Nuria:Hi all,sorry for this messy post - I forgot to subscribe to the list so I can't directly reply to your responses.
> Datasets do not include simple wiki, there are calculated for a few wikis
some or which are not very large so you might be able to use them.
> This is actually how our production search ranking is built for around the
top 20 sites by search volume that we host. Simple wikipedia isn't one of
those we currently use machine ranking for though.We use a DBN (chapelle, 2009) to transform click stream data into labeled search result data, and then LambdaMART for the final ranking model. Link to mjolnir which does the training linked above.> Because of that we do have the data you need, but the problem will be that the actual search
queries are considered PII (Personally Identifiable Information) and not
something I can release publicly. It may be possible to release aggregated
data sets that don't include the actual search terms, but at that point I
don't think the data will be useful to you anymore.I think I'm fine with query-document pairs. Isn't that sufficiently aggregated to not be considered PII?As mentioned above, the query is the hard part. Query strings contain arbitrary information and if you want to build a ranking function you have to have those original queries to do feature collection.
Just for my understanding (not a Machine Learning expert yet :) ): I
would need (query -> document) pairs such as ("machine learning" -> https://en.wikipedia.org/wiki/Machine_learning) and how often each of these pairs has ocurred, right? Even if this pair has only occured once, how is this PII? Or do I need more than just (query -> document)?
Thank you so much, this is all very enlightening!
Georg
--Thank you!Georg--Georg Sorst <g.sorst@findologic.com> schrieb am Mi., 28. Feb. 2018 um 12:17 Uhr:--Hi list,as part of a lecture on Information Retrieval I am giving we work a lot with Simple Wikipedia articles. It's a great data set because it's comprehensive and not domain specific so when building search on top of it humans can easily judge result quality, and it's still small enough to be handled by a regular computer.This year I want to cover the topic of Machine Learning for search. The idea is to look at result clicks from an internal search search engine, feed that into the Machine Learning and adjust search accordingly so that the top-clicked results actually rank best. We will be using Solr LTR for this purpose.I would love to base this on Simple Wikipedia data since it would fit well into the rest of the lecture. Unfortunately, I could not find that data. The closest I came is https://meta.wikimedia.org/wiki/Research:Wikipedia_clickstream but this covers neither Simple Wikipedia nor does it specify internal search queries.Did I miss something? Is this data available somewhere? Can I produce it myself from raw data? Ideally I would need (query-document) pairs with the number of occurrences.Thank you!GeorgGeorg M. Sorst I CTO
Jakob-Haringer-Str. 5a | 5020 Salzburg I T.: +43 662 456708
E.: g.sorst@findologic.com
www.findologic.com Folgen Sie uns auf: XING facebook TwitterWir sehen uns auf der Internet World - am 06.03. & 07.03.2018 in Halle A6 Stand E130 in München! Hier Termin vereinbaren!Wir sehen uns auf der SHOPTALK von 18. bis 21. März in Las Vegas! Hier Termin vereinbaren!Wir sehen uns auf der SOM am 18.04. & 19.04.2018 in Halle 7 Stand G.17 in Zürich! Hier Termin vereinbaren!Hier geht es zu unserer Homepage!Georg M. Sorst I CTO
Jakob-Haringer-Str. 5a | 5020 Salzburg I T.: +43 662 456708
E.: g.sorst@findologic.com
www.findologic.com Folgen Sie uns auf: XING facebook TwitterWir sehen uns auf der Internet World - am 06.03. & 07.03.2018 in Halle A6 Stand E130 in München! Hier Termin vereinbaren!Wir sehen uns auf der SHOPTALK von 18. bis 21. März in Las Vegas! Hier Termin vereinbaren!Wir sehen uns auf der SOM am 18.04. & 19.04.2018 in Halle 7 Stand G.17 in Zürich! Hier Termin vereinbaren!Hier geht es zu unserer Homepage!Georg M. Sorst I CTOFINDOLOGIC GmbH
Jakob-Haringer-Str. 5a | 5020 Salzburg I T.: +43 662 456708Wir sehen uns auf der SHOPTALK von 18. bis 21.03 in Las Vegas! Hier Termin vereinbaren!Wir sehen uns auf der SOM am 18.04. & 19.04. in Halle 7 Stand G.17 in Zürich! Hier Termin vereinbaren!Wir sehen uns auf dem SHOPWARE Community Day am 18.05. in Duisburg! Hier Termin vereinbaren!Wir sehen uns auf der OXID Commons am 14.06. in Freiburg! Hier Termin vereinbaren!Hier geht es zu unserer Homepage!______________________________________________________________________________________________
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics
