Em um pequeno refinamento no método de @eggyal, incorporei o número de itens que podemos comparar.
SELECT u2.user_id
-- join our user to their scores
FROM (users u1 JOIN scores s1 USING (user_id))
-- and then join other users and their scores
JOIN (users u2 JOIN scores s2 USING (user_id))
ON s1.item_id = s2.item_id
AND u1.user_id != u2.user_id
-- filter for our user of interest
WHERE u1.user_id = ?
-- group other users' scores together
GROUP BY u2.user_id
-- subtract the degree of difference in correlating scores from the number of correlating scores
ORDER BY (SUM(s1.item_id = s2.item_id) -
( SUM(ABS(s2.score - s1.score) + ABS(u2.self - u1.self) ) ) ) DESC
