Select equal number of random rows with respect to a column - sql

Question

Consider I have a table like this
Col1 || Col2
-------------
a || 0
b || 0
c || 1
d || 1
e || 0
How can I select rows from it so that I have equal number of 1s and 0s, like below can be a result
Col1 || Col2
-------------
a || 0
c || 1
d || 1
e || 0
The rows removed/left out are at random and deleting from an existing table would work as well.

Answer 1

For each col2 partition, you can give each row a row number and then find those rows where there is only one instance of the row number and delete them:
DELETE FROM table_name
WHERE ROWID IN (
SELECT MIN(ROWID)
FROM (
SELECT ROW_NUMBER() OVER (PARTITION BY col2 ORDER BY DBMS_RANDOM.VALUE)
AS rn
FROM table_name
)
GROUP BY rn
HAVING COUNT(*) < 2
);
If you just want to SELECT the rows then you can use a similar technique:
SELECT col1, col2
FROM (
SELECT col1,
col2,
COUNT(*) OVER (PARTITION BY rn) AS cnt
FROM (
SELECT col1,
col2,
ROW_NUMBER() OVER (PARTITION BY col2 ORDER BY DBMS_RANDOM.VALUE)
AS rn
FROM table_name
)
)
WHERE cnt = 2;
db<>fiddle here

Answer 2

How can I select rows from it so that I have equal number of 1s and 0s?
Yet another option might be to count COL2 values and use least of those two (as the final result has to have equal number of 0s and 1s) in a UNION set operation. Something like this:
Sample data:
SQL> select * from test;
COL1 COL2
---- ----------
a 0
b 0
c 1
d 1
e 0
Query & result:
SQL> with cnts as
2 -- count rows by COL2 value
3 (select sum(case when col2 = 0 then 1 else 0 end) cnt_0,
4 sum(case when col2 = 1 then 1 else 0 end) cnt_1
5 from test
6 )
7 select t.* from test t cross join cnts c
8 where t.col2 = 0 and rownum <= least(c.cnt_0, c.cnt_1)
9 union all
10 select t.* from test t cross join cnts c
11 where t.col2 = 1 and rownum <= least(c.cnt_0, c.cnt_1);
COL1 COL2
---- ----------
a 0
b 0
c 1
d 1
SQL>

Answer 3

You can do this with only one subquery/CTE. The following returns the smaller number of 0s and 1 (which determines the number of rows being returned):
least( sum(col2), sum(1 - col2) ) as num_rows
Then, you can incorporate this into a window function with row_number():
select col1, col2
from (select t.*,
least(sum(col2) over (), sum(1-col2) over ()) as num_rows,
row_number() over (partition by col2 order by dbms_random.value) as seqnum
from t
) t
where seqnum <= num_rows;

Answer 4

use the window function to count the frequency of col2 and row number over col2. Then get the minimum frequency from it. Later get the rows with rownum less than or equal to min frequency.
with data AS
(
SELECT *, row_number() over(partition by col2 order by dbms_random.value()) as rownum, COUNT(*) over(partition by col2) freq from test
),
data2 as
(
SELECT min(freq) as cnt from data
)
SELECT col1, col2 from data,data2 where rownum <= cnt

Answer 5

This analytic function check if there are more zeroes or ones in the table
sum(decode(col2,0,-1,col2)) over()
Depending on the result use cumulative sum starting with that value of col2 that appears in lower count and mapping (using decode) it to -1, the other value is mapped to 1.
The filter is done on cum_sum <= 0 i.e. you get the same number of 0 and 1.
with t1 as (
select
col1, col2,
case when sum(decode(col2,0,-1,col2)) over() <= 0 then
/* more zeroes */
sum(decode(col2,0,1,1,-1)) over(order by col2 desc, col1)
else
sum(decode(col2,0,-1,col2)) over(order by col2 , col1)
end as cum_sum
from tab)
select col1, col2
from t1
where cum_sum <= 0;