I've got four MySQL tables:
users (id, name)
polls (id, text)
options (id, poll_id, text)
responses (id, poll_id, option_id, user_id)
Given a particular poll and a particular option, I'd like to generate a table that shows which options from other polls are most strongly correlated.
Suppose this is our data set:
TABLE users:
+------+-------+
| id | name |
+------+-------+
| 1 | Abe |
| 2 | Bob |
| 3 | Che |
| 4 | Den |
+------+-------+
TABLE polls:
+------+-----------------------+
| id | text |
+------+-----------------------+
| 1 | Do you like apples? |
| 2 | What is your gender? |
| 3 | What is your height? |
| 4 | Do you like polls? |
+------+-----------------------+
TABLE options:
+------+----------+---------+
| id | poll_id | text |
+------+----------+---------+
| 1 | 1 | Yes |
| 2 | 1 | No |
| 3 | 2 | Male |
| 4 | 2 | Female |
| 5 | 3 | Short |
| 6 | 3 | Tall |
| 7 | 4 | Yes |
| 8 | 4 | No |
+------+----------+---------+
TABLE responses:
+------+----------+------------+----------+
| id | poll_id | option_id | user_id |
+------+----------+------------+----------+
| 1 | 1 | 1 | 1 |
| 2 | 1 | 2 | 2 |
| 3 | 1 | 2 | 3 |
| 4 | 1 | 2 | 4 |
| 5 | 2 | 3 | 1 |
| 6 | 2 | 3 | 2 |
| 7 | 2 | 3 | 3 |
| 8 | 2 | 4 | 4 |
| 9 | 3 | 5 | 1 |
| 10 | 3 | 6 | 2 |
| 10 | 3 | 5 | 3 |
| 10 | 3 | 6 | 4 |
| 10 | 4 | 7 | 1 |
| 10 | 4 | 7 | 2 |
| 10 | 4 | 7 | 3 |
| 10 | 4 | 7 | 4 |
+------+----------+------------+----------+
Given the poll ID 1 and the option ID 2, the generated table should be something like this:
+----------+------------+-----------------------+
| poll_id | option_id | percent_correlated |
+----------+------------+-----------------------+
| 4 | 7 | 100 |
| 2 | 3 | 66.66 |
| 3 | 6 | 66.66 |
| 2 | 4 | 33.33 |
| 3 | 5 | 33.33 |
| 4 | 8 | 0 |
+----------+------------+-----------------------+
So basically, we're identifying all of the users who responded to poll ID 1 and selected option ID 2, and we're looking through all the other polls to see what percentage of them also selected each other option.
Don't have an instance handy to test, can you see if this gets proper results:
select
poll_id,
option_id,
((psum - (sum1 * sum2 / n)) / sqrt((sum1sq - pow(sum1, 2.0) / n) * (sum2sq - pow(sum2, 2.0) / n))) AS r,
n
from
(
select
poll_id,
option_id,
SUM(score) AS sum1,
SUM(score_rev) AS sum2,
SUM(score * score) AS sum1sq,
SUM(score_rev * score_rev) AS sum2sq,
SUM(score * score_rev) AS psum,
COUNT(*) AS n
from
(
select
responses.poll_id,
responses.option_id,
CASE
WHEN user_resp.user_id IS NULL THEN SELECT 0
ELSE SELECT 1
END CASE as score,
CASE
WHEN user_resp.user_id IS NULL THEN SELECT 1
ELSE SELECT 0
END CASE as score_rev,
from responses left outer join
(
select
user_id
from
responses
where
poll_id = 1 and
option_id = 2
)user_resp
ON (user_resp.user_id = responses.user_id)
) temp1
group by
poll_id,
option_id
)components
After a few hours of trial and error, I managed to put together a query that works correctly:
SELECT poll_id AS p_id,
option_id AS o_id,
COUNT(*) AS optCount,
(SELECT COUNT(*) FROM response WHERE option_id = o_id AND user_id IN
(SELECT user_id FROM response WHERE poll_id = '1' AND option_id = '2')) /
(SELECT COUNT(*) FROM response WHERE poll_id = p_id AND user_id IN
(SELECT user_id FROM response WHERE poll_id = '1' AND option_id = '2'))
AS percentage
FROM response
INNER JOIN
(SELECT user_id FROM response WHERE poll_id = '1' AND option_id = '2') AS user_ids
ON response.user_id = user_ids.user_id
WHERE poll_id != '1'
GROUP BY option_id DESC
ORDER BY percentage DESC, optCount DESC
Based on a tests with a small data set, this query looks to be reasonably fast, but I'd like to modify it so the "IN" subquery is not repeated three times. Any suggestions?
This seems to give the right results for me:
select poll_stats.poll_id,
option_stats.option_id,
(100 * option_responses / poll_responses) as percent_correlated
from (select response.poll_id,
count(*) as poll_responses
from response selecting_response
join response on response.user_id = selecting_response.user_id
where selecting_response.poll_id = 1 and selecting_response.option_id = 2
group by response.poll_id) poll_stats
join (select options.poll_id,
options.id as option_id,
count(response.id) as option_responses
from options
left join response on response.poll_id = options.poll_id
and response.option_id = options.id
and exists (
select 1 from response selecting_response
where selecting_response.user_id = response.user_id
and selecting_response.poll_id = 1
and selecting_response.option_id = 2)
group by options.poll_id, options.id
) as option_stats
on option_stats.poll_id = poll_stats.poll_id
where poll_stats.poll_id <> 1
order by 3 desc, option_responses desc
来源：https://stackoverflow.com/questions/5312166/advanced-mysql-find-correlations-between-poll-responses