SQL- Sample the 3rd from the beginning and backwards - sql

Question

I have a table with id and score. I want to create a new set of data with a sampling method. The sampling method would be to order the id in decreasing order of the scores and sample the 3rd id, starting with the first form the beginning until we get 10k positive samples. And we would like to do the same in the other direction, starting from the end to get 10k negative samples.
id
score
24
0.55
58
0.43
987
0.93
How can I write a SQL query to execute this sampling and get the expected output?

Answer 1

To start with, this would be more straightforward to write an answer if you included the database you used (SQL Server, MySQL, etc). Different SQL versions have different syntax.
BACKGROUND
To answer this question, the main tools you need are the ability to sort, and an ability to take every 3rd row.
I'm using SQL Server here, so sorting includes
TOP modifier in SELECT statements - in other databases it's often LIMIT (e.g., SELECT * FROM Test LIMIT 1000)
ROW_NUMBER() which I believe is relatively common
To get every third row, I use the 'modulus' mathematical function - in SQL Server signified by a % symbol - so, for example
1 % 3 = 1
2 % 3 = 2
3 % 3 = 0
4 % 3 = 1
APPROACH
There is an example of this in this db<>fiddle - but note that it is only dealing with test data (1000 rows, selecting top and bottom 10).
Running through the steps - and assuming your data is stored in #DataTable:
The following command assigns a row number rn to the data, sorted by the score.
SELECT id, score, ROW_NUMBER() OVER (ORDER BY score, id) as rn
FROM #DataTable;
To get every third value, start with that data and take every third value (e.g., where the row number is a multiple of 3)
SELECT *
FROM (SELECT id, score, ROW_NUMBER() OVER (ORDER BY score, id) as rn
FROM #DataTable)
WHERE rn % 3 = 0;
To get the first 10,000 of them, use TOP (or LIMIT, etc)
SELECT TOP 10000 *
FROM (SELECT id, score, ROW_NUMBER() OVER (ORDER BY score, id) as rn
FROM #DataTable)
WHERE rn % 3 = 0
ORDER BY rn;
Note - to get it the other way/get the highest scores, take the ROW_NUMBER() in reverse order (e.g., ORDER BY score DESC, id DESC).
FINAL ANSWER
Take the above 10,000 rows, and do a similar for the other way (e.g., to get the highest scores) then UNION them together. Below it is done with a CTE.
WITH TopScores AS
(SELECT TOP 10000 id, score
FROM (SELECT id, score, ROW_NUMBER() OVER (ORDER BY score DESC, id DESC) as rn
FROM #DataTable
) AS RankedScores_down
WHERE RankedScores_down.rn % 3 = 0
ORDER BY RankedScores_down.rn
),
LowScores AS
(SELECT TOP 10000 id, score
FROM (SELECT id, score, ROW_NUMBER() OVER (ORDER BY score, id) as rn
FROM #DataTable
) AS RankedScores_up
WHERE RankedScores_up.rn % 3 = 0
ORDER BY RankedScores_up.rn
)
SELECT * FROM TopScores
UNION
SELECT * FROM LowScores
ORDER BY score, id;
Notes
I used 'UNION' rather than 'UNION ALL' because, in the chance that there is overlap (e.g., you have less than 60,000 datapoints) we only want to include each sample once
If you use a different database, you'll need to translate this! Here are the benefits of specifying the database you use.
Note that taking every third value (when sorted by score) is not really 'independent' sampling - one would ask why you just don't use all of the top/bottom 30,000 scores? If you to sample 1 in 3 of them, instead you could use id % 3 instead of rn % 3. But once again, why would you do this? Why not just collect fewer results and use them all?
Of course, one good reason is to use half the data to check the validity of stats e.g., take half the data, do your model - then check against the other half how good your model is.