ENH/API: DataFrame's isin should accept DataFrames

[TomAugspurger]

From this SO question.

The user had two dataframes and wanted to use the second as the values argument to isin.

We'd need to think about how to handle the other index. In this case the user only cared about the columns, not the index labels.

Previous issues/PRs: #4258 and #4211

[hayd]

So, as you say, at the moment this works by passing in:

df1.isin(df2.to_dict(outtype='list'))

Is this an efficient operation? Should we just call that if passed a DataFrame?

[jreback]

FYI I believe this will fail with non-unique columns (as to_dict will fail); look at itertuples on a way to deal with this

[jreback]

Should isin accept something like a set operator mode?

maybe mode=diff for set differences, or mode=and for inclusion?

http://stackoverflow.com/questions/18180763/set-difference-for-pandas/18187648#18187648

[hayd]

see #4617 from same OP.

I'm not sure I'm 100% with how this was going to work, because the column approach no longer makes sense once we're talking about DataFrames, or at least is more ambiguous (aside from repeated columns):

In [2]: df1 = pd.DataFrame([[1, 2], [2, 3]])

In [3]: df2 = pd.DataFrame([[1, 3]])

In [4]: df2.isin(df1.to_dict(outtype='list'))
Out[4]: 
      0     1
0  True  True

which probably isn't what the user was expecting.

Which may be something like (possibly user doesn't care about the index):

In [11]: pd.Series(map(set(df1.itertuples()).__contains__, df2.itertuples()), df2.index)
Out[11]: 
0    False
dtype: bool

Ahem... Not sure what would be an efficient way to do that?

@jreback imo set intersection stuff can relatively easily be done after the fact, e.g. (all/any and boolean indexing, really don't think we need to add loads of kwargs.

[jreback]

@hayd @TomAugspurger for 0.13?

[hayd]

Definitely.

[jreback]

@hayd @TomAugspurger what is left on this?

[hayd]

To actually write this up... I think it's kind of important for 0.13 as otherwise isin is a little confusing (and will break code if we change the API later)....

Thinking about it again, do we actually want to eq (ag. there a problem with dupes here...):

In [32]: df1 = pd.DataFrame([[1, 2], [2, 3]], columns=['a', 'b'])

In [33]: df2 = pd.DataFrame([[1, 3]], columns=['a', 'b'])

In [34]: df2.isin(df1.to_dict(outtype='list'))
Out[34]: 
      a     b
0  True  True

In [35]: df2.eq(df1.reindex_like(df2))  # is this isin for DataFrames?
Out[35]: 
      a      b
0  True  False

Worst case we should NotImplement it with a mention of eq...

[TomAugspurger]

@hayd Are you working on this one? I can give it a shot tonight / tomorrow if you're wanting to focus on the SQLAlchemy stuff.

---

Let's collect some thoughts on what should happen when another DataFrame is passed as values.

1. Identical indicies, no dupes

In [20]: df1
Out[20]: 
   A   B
0  1   2
1  2 NaN
2  3   4
3  4   4

In [21]: df2
Out[21]: 
    A   B
0   0   2
1   2 NaN
2  12   4
3   4   5

# Expected: df1.isin(df2.to_dict(outtype='list')) does not work
# df1.eq(df2.reindex_like(df1)) does work
Out[23]: 
       A      B
0  False   True
1   True  False
2  False   True
3   True   False

1. Dissimilar indices, still no dupes. Should be intersection of indices, check isin on those (optional parameter to ignore index?).

# df1 same as before
In [35]: df2 = pd.DataFrame({'A': [0, 2, 12, 4], 'C': [2, np.nan, 4, 5]}, index=[0, 1, 3, 4])

In [36]: df2
Out[36]: 
    A   C
0   0   2
1   2 NaN
3  12   4
4   4   5

# Intersection of columns is A, intersection of indices is [0, 1, 3]
# df1.isin(df2.to_dict(outtype='list')) does *NOT* work since it matches (A, 3), it ignores the index.
# Andy's df1.eq(df2.reindex_like(df1)) does work here.
In [39]: df1.isin(df2.to_dict(outtype='list'))
Out[39]: 
       A      B
0  False  False
1   True  False
2  False  False
3  False  False

1. Duplicate columns on values (the argument). The two options are to raise or return True wherever there's a match.

In [74]: df1
Out[74]: 
   A   B
0  1   2
1  2 NaN
2  3   4
3  4   4

In [75]: df2
Out[75]: 
    A   A
0   0   1
1   2   4
3   2 NaN
4   4   5

# Expected if we return True wherever there's a True

       A      B
0   True  False   # from the second A column
1   True  False   # from the first A column
2  False  False
3  False  False

The df1.eq(df2.reindex_like(df1)) does not work here, this hits the maximum recursion depth.
I had my dfs backwards; that's actually a ValueError: cannot reindex from a duplicate axis.

[TomAugspurger]

Oh we definitely need to get this in for .13. I think DataFrames and Series should behave similarly with respect to labels. Currently, if values is a Series, the index is ignored:

In [150]: df = pd.DataFrame({'A': [1, 2, 3, 4], 'B': [2, np.nan, 4, 4]}, index=['a','b','c','d'])

In [151]: s = pd.Series([1, 3, 11, 12], index=['a','b','c','d'])

In [152]: df.isin(s)
Out[152]: 
       A      B
a   True  False
b  False  False
c   True  False
d  False  False

So (A, c) matches on the 3 from the Series, despite the index of that 3 being b, not c.

What's everyone's thoughts on respecting index/column labels? I'd say that (A, c) should be False here since the labels don't match.

[jreback]

@TomAugspurger there was a bug in comparing duplicate frames (so your above comparison will work as you had it, or raise if you don't reindex_like), but won't infinity recurse (which generally is a bad thing :)

will merge in a few

In [1]: df1 = DataFrame([[1,2],[2,np.nan],[3,4],[4,4]],columns=['A','B'])

In [2]: df2 = DataFrame([[0,1],[2,4],[2,np.nan],[4,5]],columns=['A','A'])

In [3]: df1
Out[3]: 
   A   B
0  1   2
1  2 NaN
2  3   4
3  4   4

In [4]: df2
Out[4]: 
   A   A
0  0   1
1  2   4
2  2 NaN
3  4   5

In [5]: df1.reindex_like(df2) == df2
Out[5]: 
       A      A
0  False   True
1   True  False
2  False  False
3   True  False

[jreback]

@TomAugspurger if you need to test if the the columns are duplicated, use df.columns.is_unique