Well, sort of "different" (I'd never assume that the names meant "join" myself)...

But I think this does it.

The big inline view pivots the primary key columns for all tables in a schema.

Then we join this to every other table in the schema (cartesian join...) and for each table -- see if that other table has all of those columns.

I'm not going to say "it'll be fast", but it should work ;)

The NO_MERGE may or may not be needed depending on an "optimizer" issue (if you get a group by error, you need it)

select p.table_name, c.table_name
from (
select /*+ NO_MERGE */ table_name,
max(decode(rn,1,column_name)) c1,
max(decode(rn,2,column_name)) c2,
max(decode(rn,3,column_name)) c3,
max(decode(rn,4,column_name)) c4,
max(decode(rn,5,column_name)) c5,
max(decode(rn,6,column_name)) c6,
max(decode(rn,7,column_name)) c7,
max(decode(rn,8,column_name)) c8,
count(*) cnt
from (
select a.table_name, b.column_name,
row_number() over (partition by a.table_name order by b.column_name) rn
from user_constraints a,
user_cons_columns b
where a.constraint_type = 'P'
and a.constraint_name= b.constraint_name
)
group by table_name
) P,
user_tables C
where p.table_name <> c.table_name
and (select count(*)
from user_tab_columns
where column_name in ( p.c1, p.c2, p.c3, p.c4, p.c5, p.c6, p.c7, p.c8 )
and table_name = C.table_name
) = p.cnt
/