Gregory Todd Williams

I recently began taking a look at the Wikidata query logs that were published a couple of months ago and wanted to look into how some features of SPARQL were being used on Wikidata. The first thing I’ve looked at is the use of property paths: how often paths are used, what path operators are used, and with what frequency.

Using the “interval 3” logs (2017-08-07–2017-09-03 representing ~78M successful queries¹), I found that ~25% of queries used property paths. The vast majority of these use just a single property path, but there are queries that use as many as 19 property paths:

Pct.	Count	Number of Paths
74.3048%	58161337	0 paths used in query
24.7023%	19335490	1 paths used in query
0.6729%	526673	2 paths used in query
0.2787%	218186	4 paths used in query
0.0255%	19965	3 paths used in query
0.0056%	4387	7 paths used in query
0.0037%	2865	8 paths used in query
0.0030%	2327	9 paths used in query
0.0011%	865	6 paths used in query
0.0008%	604	11 paths used in query
0.0006%	434	5 paths used in query
0.0005%	398	10 paths used in query
0.0002%	156	12 paths used in query
0.0001%	110	15 paths used in query
0.0001%	101	19 paths used in query
0.0001%	56	13 paths used in query
0.0000%	12	14 paths used in query

I normalized IRIs and variable names used in the paths so that I could look at just the path operators and the structure of the paths. The type of path operators used skews heavily towards * (ZeroOrMore) as well as sequence and inverse paths that can be rewritten as simple BGPs. Here are the structures representing at least 0.1% of the paths in the dataset:

Pct.	Count	Path Structure
49.3632%	10573772	?s <iri1> * ?o .
39.8349%	8532772	?s <iri1> / <iri2> ?o .
4.6857%	1003694	?s <iri1> / ( <iri2> * ) ?o .
1.8983%	406616	?s ( <iri1> + ) / ( <iri2> * ) ?o .
1.4626%	313290	?s ( <iri1> * ) / <iri2> ?o .
1.1970%	256401	?s ( ^ <iri1> ) / ( <iri2> * ) ?o .
0.7339%	157212	?s <iri1> + ?o .
0.1919%	41110	?s ( <iri1> / ( <iri2> * ) ) / ( ^ <iri3> ) ?o .
0.1658%	35525	?s <iri1> / <iri2> / <iri3> ?o .
0.1496%	32035	?s <iri1> / ( <iri1> * ) ?o .
0.1124%	11889	?s ( <iri1> / <iri2> ) / ( <iri3> * ) ?o .

There are also some rare but interesting uses of property paths in these logs:

Pct.	Count	Path Structure
0.0499%	5274	?s ( ( <iri1> / ( <iri2> * ) ) / ( <iri3> / ( <iri2> * ) ) ) / ( <iri4> / ( <iri2> * ) ) ?o .
0.0015%	157	?s ( <iri1> / <iri2> / <iri3> / <iri4> / <iri5> / <iri6> / <iri7> / <iri8> / <iri9> ) * ?o .
0.0003%	28	?s ( ( ( ( <iri1> / <iri2> / <iri3> ) ? ) / ( <iri4> ? ) ) / ( <iri5> * ) ) / ( <iri6> / ( <iri7> ? ) ) ?o .

Without further investigation it’s hard to say if these represent meaningful queries or are just someone playing with SPARQL and/or Wikidata, but I found them curious.