Add set operations to Series objects

[cpcloud]

from this SO question

[hayd]

Would this just be on the Series values (i.e. ignore index)?

I had a look in algos and couldn't see anything (np intersect1d seems slow).

OT but a weird thing from that Q is that you can't call Series on a set.

[cpcloud]

yet a frozenset works....werid...i would've though that frozenset was a subclass of set, guess not

[cpcloud]

well it doesn't work...it returns a frozenset...

[hayd]

(I don't see why we enforce this, list is happy to take a set, why shouldn't Series?)

[cpcloud]

because there's no way to map indices to set, they are arbitrary since a set object is unordered

[hayd]

ha! Series let's a lot of stuff drop through.... eg. Series(1)

[cpcloud]

but you're right....list does it so there must be some arbitrary indices assigned

[cpcloud]

i just discovered that Series takes generators! i had no idea.

[hayd]

and that's the workaround, right? pass it to list first... kinda sucks Also isn't a dict similarly unordered (and yet we allow that) :s

Interestingly I thought we used np.fromiter to do that, but apparently it's just list.

[cpcloud]

but dict has keys which are used as the index

[cpcloud]

i guess it's a workaround, but how else would you do it?

[cpcloud]

i made the change....nothing breaks...i'll submit

[cpcloud]

for set construction that is

[cpcloud]

#4482

[jreback]

fyi...the Series(1) stuff is getting squash in #3862 (by #3482), its just odd

[hayd]

Cool bananas

[cpcloud]

won't be adding this, so closing

[ghost]

@cpcloud, just making sure your closing this is unrelated to the discussion in #4482,
which was only about modifying the series ctor to accept sets, not about adding
set operations to series.

[cpcloud]

you're right...that pr was just to disallow frozenset and was only related to the ctor...reopening

[jreback]

@cpcloud what's the status on this?

[cpcloud]

gone by wayside .... i don't really have time to implement this ... but i think we should leave it open ... marking as someday

[jreback]

ok..gr8 thxs

[jreback]

most of these I just push to 0.14.....someday is a box very rarely opened :)

[makmanalp]

Hmm, is this closed now since 0.14 is out?

[jreback]

@makmanalp what are you trying to do?

[makmanalp]

Efficiently calculate which of the rows of a column in df1 also exist in another column in df2 (or perhaps indices instead of columns).

[jreback]

this issue is a bit different than that (see the linked question), did you try isin?

[hayd]

In the original question it looks like the OP wants to ignore the index... in which case they can use the set operations in Index:

pd.Index(s0.values) & pd.Index(s1.values)

[hayd]

Wow, that's really slow, I take that back...

[jreback]

Index keeps things ordered; should'nt do it that way, better to drop into numpy or python, do the set operation and reconstruct.

[milindsmart]

Any update on this? Is this still impossible? I was looking at doing a symmetric difference.

[h-vetinari]

This is something I would have needed several times in the last half year. Using apply is just terribly slow for large data sets (and numpy has fast set implementations like np.intersect1d) - for example, I have code where apply+intersect is 98% of running time.

To recap (since there's a lot of tangential discussion in this thread), I think there is a good case to be made for a .set-accessor, providing access to universal functions operating on sets, like there currently is for .str, .dt. As an example:

import pandas as pd # 0.21.0
import numpy as np # 1.13.3

# this function is just for demonstration purposes
def random_sets(n = 100):
    length = 10
    # strings of random numbers, padded to common length
    s = pd.Series(np.random.randint(0, 10**length-1, (n,), dtype = np.int64)).astype(str).str.zfill(length)
    # split into set of individual numbers and cast back to integers
    return s.map(set).apply(lambda s: set(int(x) for x in s)) 

a = random_sets(5)
# WANTED: a.set.intersect(set([1, 2])) should result in:
a.apply(lambda r: r & set([1, 2])) # intersection of 'a' with {1, 2}, e. g.
# 0    {1, 2}
# 1       {1}
# 2        {}
# 3    {1, 2}
# 4        {}

b = random_sets(5)
# WANTED: a.set.intersect(b) should result in:
pd.concat([a, b], keys = ['a', 'b'], axis = 1).apply(lambda row: row['a'] & row['b'], axis = 1) # intersection of 'a' and 'b' (per row!), e. g.
# 0          {8, 1, 2}
# 1    {1, 3, 5, 7, 9}
# 2       {0, 8, 5, 9}
# 3       {8, 2, 3, 6}
# 4       {8, 9, 3, 6}

Like the .str-methods, it should work with either another Series (including index alignment), or broadcast a 'scalar' set correspondingly. Some important methods I think should be implemented (the function signature tries to indicate the action per row; the names are just suggestions):

(set, set) -> set:
a.set.intersect(b) - as above
a.set.union(b) - row-wise union of a and b
a.set.diff(b) - row-wise set difference of a and b
a.set.xor(b) - row-wise symmetric difference of a and b

(set, set) -> bool:
a.set.subset(b) - row-wise check if a is subset of b
a.set.superset(b) - row-wise check if a is superset of b

(set, obj) -> bool:
a.set.contains(c) - row-wise check if a contains c

[jreback]

@h-vetinari sets are not efficiently stored, so this offers only an api benefit, which I have yet to see and interesting use case. You can use Index for these operations individually and that is quite efficient. Series.isin is pretty much .contains. IntervalIndex will work for some of these cases as well.

[h-vetinari]

@jreback, well, I was hoping not just for an API improvement, but some fast cython code to back it up (like for the .str-methods). ;-)

Do I understand you correctly that you propose to work with Series of (short) pd.Indexes? How would you then do something like a.set.intersect(b) (as described above)?

[jreback]

@h-vetinari and you are welcome to contribute things. I the current impl would be quite inefficient and no each way to get around this ATM.

[chinchillaLiao]

import pandas as pd
df = pd.DataFrame({'a':{1,2,3}, 'b':{2,3,4}})

Difference operator works between Series.

df['a - b'] = df['a'] - df['b']

Set intersection operactor doesn't work between Series.

df['a & b'] = df['a'] & df['b']

A very slow way to do intersection between Series:

df['a & b'] = df.apply(lambda row: row['a'] & row['b'], axis = 1)

I found it is much more faster to do intersection this way:

df['a & b'] = df['a'] - (df['a'] - df['b'])

I don't know why.

[h-vetinari]

@chinchillaLiao : cool, didn't know set difference worked on Series! It's the only one to work on pandas level though.

But an even better work-around is to go down to the numpy-implementation with .values. In particular, this shouldn't suffer from the speed degradation you're reporting.

(@jreback; my comment half a year ago about a .set-accessor now seems very superfluous - why not just enable the numpy-behaviour directly in pandas?)

df = pd.DataFrame([[{1,2}, {2,3}],[{2,4}, {3, 1}]], columns=['A', 'B'])
df
#         A       B
# 0  {1, 2}  {2, 3}
# 1  {2, 4}  {1, 3}

df['A - B'] = df.A - df.B  # only one that work out of the box in pandas
df['A - B']
# 0       {1}
# 1    {2, 4}
# dtype: object

df['A & B'] = df.A & df.B
# TypeError: unsupported operand type(s) for &: 'set' and 'bool'
df['A & B'] = df.A.values & df.B.values
df['A & B'] 
# 0    {2}
# 1     {}
# Name: A & B, dtype: object

df['A | B'] = df.A | df.B
# TypeError: unsupported operand type(s) for |: 'set' and 'bool'
df['A | B'] = df.A.values | df.B.values
df['A | B'] 
# 0       {1, 2, 3}
# 1    {1, 2, 3, 4}
# Name: A | B, dtype: object

df['A ^ B'] = df.A ^ df.B
# TypeError: unsupported operand type(s) for ^: 'set' and 'bool'
df['A ^ B'] = df.A.values ^ df.B.values
df['A ^ B'] 
# 0          {1, 3}
# 1    {1, 2, 3, 4}
# Name: A ^ B, dtype: object

df
#         A       B   A - B A & B         A | B         A ^ B
# 0  {1, 2}  {2, 3}     {1}   {2}     {1, 2, 3}        {1, 3}
# 1  {2, 4}  {1, 3}  {2, 4}    {}  {1, 2, 3, 4}  {1, 2, 3, 4}

In terms of usability, the really cool thing is that this also works for many-to-one comparisons.

dd = df.A.to_frame()
C = {2, 5}
dd['A - C'] = df.A - C
dd['A & C'] = df.A.values & C
dd['A | C'] = df.A.values | C
dd['A ^ C'] = df.A.values ^ C

dd
#         A A - C A & C      A | C   A ^ C
# 0  {1, 2}   {1}   {2}  {1, 2, 5}  {1, 5}
# 1  {2, 4}   {4}   {2}  {2, 4, 5}  {4, 5}

[jreback]

sets are not first class and actually completely inefficient in a Series

[h-vetinari]

Inefficient as opposed to what? Some situations fundamentally require processing sets.

And even so, why make treating sets harder than it needs to be? I used to think (see my response from December) that this wasn't implemented at all, but since it's in numpy already, why not just expose that functionality on a pandas level? Sure as hell beats writing your own .apply() loops, both in terms of speed and code complexity.

[jreback]

complexity in terms of implementation and code

sure if u wanted to contribute would be great

but it’s not trivial to do in a first class supoorted way

[hayd]

@h-vetinari what is your use case for this? How does this come up?

IMO a nice way to contribute this would be with an extension type in a library. Depending on your use case. If you have a smallish finite super-set you can describe each set as a bitarray (and hence do set operations cheaply).

Note: This is quite different from the original issue: set operations like set(s1) & set(s2)...

[h-vetinari]

@jreback

OK, I'm thinking about contributing that. Since the numpy-methods I showed above are actually not nan-safe,

np.array([{1,2}, np.nan]) | np.array([{2,4}, {3, 1}])
# TypeError: unsupported operand type(s) for |: 'float' and 'set'

I'm back to thinking that a set accessor for Series (not for Index) would be the best. And, since I wouldn't have to write the cython for those methods, I think I can come up with such a wrapper relatively easily.

@hayd
Thanks for the link. I've had several use cases over time, e.g. joining email-addresses/telephone numbers when deduplicating user information. But it keeps cropping up. I'm actually really happy I found out about those numpy-methods some days ago. ;-)

Re:

Note: This is quite different from the original issue: set operations like set(s1) & set(s2)...

I've chosen to comment on this issue (rather than opening a new one) due to the title, which imo has a much larger scope. I could easily open a more general issue, if desired.

[jbrockmendel]

Discussed on today's dev call and the consensus was to convert your Series to Index and do setops on those. Closing.