We already to have an existing recommender system. It should be improved. It is very important for me that we do not hard-code classes for the recommendation.

Great. I figured that was the case. What would be the best way to understand how prediction already works and to find out where improvements could be made?

Related: Steinhauser, S., Gassler, W., Pichl, M., & Zangerle, E. Evaluation of Property Recommenders for Wikidata. http://informatik.uibk.ac.at/en/teaching/smb/theses/410.pdf

See also: Zangerle, E., Gassler, W., Pichl, M., Steinhauser, S., & Specht, G. (2016, August). An Empirical Evaluation of Property Recommender Systems for Wikidata and Collaborative Knowledge Bases. In Proceedings of the 12th International Symposium on Open Collaboration (p. 18). ACM. http://www.opensym.org/os2016/proceedings-files/p503-zangerle.pdf

@Lydia_Pintscher, is this the docs for the current recommender? https://www.mediawiki.org/wiki/WikidataEntitySuggester

https://www.wikidata.org/wiki/Property:P1963

In T156401#3007483, @Halfak wrote:

@Lydia_Pintscher, is this the docs for the current recommender? https://www.mediawiki.org/wiki/WikidataEntitySuggester

Yes.

If work on this is started let's talk about it first.

@Lydia_Pintscher just showed me https://cool-wd.inf.unibz.it/ and apparently there's also a gadget for recommending properties from the same group.

Here's the gadget https://www.wikidata.org/wiki/User:Aahmeti/recool-wd.js

FYI, I talked to @Sjoerddebruin about this at WikiCite and he showed me https://www.wikidata.org/wiki/User:Sjoerddebruin/Entity_suggester I think that we should consider this when trying to make improvements.

I wonder if blaze graph this. @Smalyshev, how familiar are you with the property suggestion system and if blaze graph might be a good solution for recommend properties for entities (items, properties themselves)?

@Halfak depends on whether we can make a query that is fast enough. I think @Jonas did some work on that and for some cases queries are fast, while for others they are rather slow and it would better work as pre-calculated data set.