Getting Random Results in SPARQL

Just the other day I decided to take a stab at an old StackOverflow questing about getting random results from SPARQL.

The most obvious solution might first appear to be using SPARQL’s built-in RAND() function and order by that random number:

SELECT ?s WHERE {
  ?s ?p ?o .
  BIND(RAND() AS ?random) .
} ORDER BY ?random
LIMIT 1

This could have been perfectly fine if it weren’t for SPARQL engines trying to be smart and statically evaluating the third line. At the second result row, most SPARQL engines see a line and “thinks”, “oh this is identical to what I did on the row before this one, the result must be the same”.

This can be illustrated by the Wikidata SPARQL query below, note how all rows have the same ?random value:

SELECT ?item ?itemLabel ?random WHERE {
  ?item wdt:P31 wd:Q11762356 .
  BIND(RAND() AS ?random) .
  SERVICE wikibase:label { bd:serviceParam wikibase:language "[AUTO_LANGUAGE], en" . }
} ORDER BY ?random
LIMIT 100
# comment, change this before each run to bypass WDQS cache

A common solution to this issue is ditch RAND() entirely and instead hash a value in each row and sort it by the hash.

SELECT ?s WHERE {
  ?s ?p ?o .
  BIND(MD5(?s) AS ?random) .
} ORDER BY ?random
LIMIT 1

This will however generate the same order each time as the hashes are entirely based on the result data. Illustrated by the example below:

SELECT ?item ?itemLabel ?random WHERE {
  ?item wdt:P31 wd:Q11762356 .
  BIND(MD5(STR(?item)) AS ?random) .
  SERVICE wikibase:label { bd:serviceParam wikibase:language "[AUTO_LANGUAGE], en" . }
} ORDER BY ?random
LIMIT 100
# comment, change this before each run to bypass WDQS cache

The solution is to combine the two. The use of hashed result data makes sure the SPARQL engine can’t statically evaluate RAND() and the fact that RAND() is therefore executed each time helps avoid the selection bias.

SELECT ?s WHERE {
  ?s ?p ?o .
  BIND(SHA512(CONCAT(STR(RAND()), STR(?s))) AS ?random) .
} ORDER BY ?random
LIMIT 1

Here again illustrated with a Wikidata query:

SELECT ?item ?itemLabel ?random WHERE {
  ?item wdt:P31 wd:Q11762356 .
  BIND(SHA512(CONCAT(STR(RAND()), STR(?item))) AS ?random) .
  SERVICE wikibase:label { bd:serviceParam wikibase:language "[AUTO_LANGUAGE], en" . }
} ORDER BY ?random
LIMIT 100
# comment, change this before each run to bypass WDQS cache

Know a better way to retrieve random results from SPARQL? Let me know on Twitter!