Window function inside a pivot, is it possible?

I have this table

And I need to create a pivot, that displays emotions as columns, with average emotion_level grouped by user_id, user_date, emotion. For example, for user_id = 1, user_date = 2011-07-13 and emotion = 'Anger', the average emotion_level should be 4.0.

I create a pivot:

select USER_ID, user_date,  
AVG(case emotion when 'Anger' then convert(float, emotion_level) else 0 end)  as Anger,
AVG(case emotion when 'Sadness' then convert(float, emotion_level) else 0 end) as Sadness,
AVG(case emotion when 'Interest' then convert(float, emotion_level) else 0 end) as Interest
from emotions group by USER_ID, user_date;

Which half-works, but calculates average emotion_level among all emotions, but not for emotions grouped by user, date and emotion.

My result for first user + emotion = 'Anger' = 2,开发者_如何学Python but it should be 4.

I guess, I should use the window function (over (partition by user_id, user_date, emotion)), but can't get the syntax run.

Is it possible at all?

I'm using PostgreSQL 9 in prod, but the above example is written in SQL Server.

WITH    q (id, user_id, user_date, emotion, emotion_level) AS
        (
        VALUES
        (1, 1, '2011-07-13'::DATE, 'Anger', 3),
        (2, 1, '2011-07-13'::DATE, 'Anger', 5),
        (3, 1, '2011-07-13'::DATE, 'Sadness', 2),
        (4, 1, '2011-07-13'::DATE, 'Interest', 2),
        (5, 2, '2011-07-13'::DATE, 'Anger', 1),
        (6, 2, '2011-07-13'::DATE, 'Sadness', 4),
        (7, 2, '2011-07-13'::DATE, 'Sadness', 5),
        (8, 2, '2011-07-13'::DATE, 'Interest', 3),
        (9, 3, '2011-07-13'::DATE, 'Anger', 1),
        (10, 3, '2011-07-13'::DATE, 'Sadness', 3),
        (11, 3, '2011-07-13'::DATE, 'Interest', 4),
        (12, 3, '2011-07-13'::DATE, 'Interest', 5)
        ) 
SELECT  user_id, user_date,
        AVG(CASE emotion WHEN 'Anger' THEN emotion_level END)::numeric(3, 2) AS Anger,
        AVG(CASE emotion WHEN 'Sadness' THEN emotion_level END)::numeric(3, 2) AS Sadness,
        AVG(CASE emotion WHEN 'Interest' THEN emotion_level END)::numeric(3, 2) AS Interest
FROM    q
GROUP BY
        user_id, user_date
ORDER BY
        user_id, user_date

The problem was that the expressions you originally used:

AVG(case emotion when 'Interest' then convert(float, emotion_level) else 0 end)

averaged over all records for a given user at given date and treated non-Interest entries as 0, while they should be treated as NULL so they would not contribute to Interest average.

I first updated the datatype of 'emotion_level' to 'float' in the table and then performed the operation using pivot. If the datatype is not updated, the required values are of int datatype.

select user_id, user_date, Anger, Sadness, Interest 
from (select user_id,user_date,emotion, emotion_level 
from emotions) as emo
pivot(avg(emotion_level) 
for emotion in (Anger, Sadness, Interest)) as P;

The result from the above query in the link below

I'd define a CTE, then join to it - your main problem is probably the fact that you're not grouping on emotion_level. Try this:

WITH average_emotion (user_id, date, emotion, average_level) as 
                     (SELECT user_id, user_date, emotion, AVG(convert(float, emotion_level))
                      FROM emotions
                      GROUP BY user_id, user_date, emotion)
SELECT a.user_id, a.user_date, COALESCE(b.average_level, 0) as Anger
FROM emotions as a
LEFT JOIN average_emotion as b
ON b.user_id = a.user_id
AND b.date = a.user_date
AND b.emotion = 'Anger'
GROUP BY a.user_id, a.user_date

Then, just add additional left joins every time you need a new column.

There's a number of ways to restructure this, mostly relating to the rest of your data model, and desired output.