How to Query Multiindex Index Columns Values in Pandas

How to query MultiIndex index columns values in pandas

To query the df by the MultiIndex values, for example where (A > 1.7) and (B < 666):

In [536]: result_df = df.loc[(df.index.get_level_values('A') > 1.7) & (df.index.get_level_values('B') < 666)]

In [537]: result_df
Out[537]: 
          C
A   B      
3.3 222  43
    333  59
5.5 333  56

Hence, to get for example the 'A' index values, if still required:

In [538]: result_df.index.get_level_values('A')
Out[538]: Index([3.3, 3.3, 5.5], dtype=object)

The problem is, that in large data frames the performance of by index selection worse by 10% than the sorted regular rows selection. And in repetitive work, looping, the delay accumulated. See example:

In [558]: df = store.select(STORE_EXTENT_BURSTS_DF_KEY)

In [559]: len(df)
Out[559]: 12857

In [560]: df.sort(inplace=True)

In [561]: df_without_index = df.reset_index()

In [562]: %timeit df.loc[(df.index.get_level_values('END_TIME') > 358200) & (df.index.get_level_values('START_TIME') < 361680)]
1000 loops, best of 3: 562 µs per loop

In [563]: %timeit df_without_index[(df_without_index.END_TIME > 358200) & (df_without_index.START_TIME < 361680)]
1000 loops, best of 3: 507 µs per loop

Python/Pandas - Query a MultiIndex Column

To answer my own question, it looks like query shouldn't be used at all (regardless of using MultiIndex columns) for selecting certain columns, based on the answer(s) here:

Select columns using pandas dataframe.query()

Select rows in pandas MultiIndex DataFrame

MultiIndex / Advanced Indexing

Note

This post will be structured in the following manner:
The questions put forth in the OP will be addressed, one by one
For each question, one or more methods applicable to solving this problem and getting the expected result will be demonstrated.
Notes (much like this one) will be included for readers interested in learning about additional functionality, implementation details,
and other info cursory to the topic at hand. These notes have been
compiled through scouring the docs and uncovering various obscure
features, and from my own (admittedly limited) experience.
All code samples have created and tested on pandas v0.23.4, python3.7. If something is not clear, or factually incorrect, or if you did not
find a solution applicable to your use case, please feel free to
suggest an edit, request clarification in the comments, or open a new
question, ....as applicable.

Here is an introduction to some common idioms (henceforth referred to as the Four Idioms) we will be frequently re-visiting

DataFrame.loc - A general solution for selection by label (+ pd.IndexSlice for more complex applications involving slices)
DataFrame.xs - Extract a particular cross section from a Series/DataFrame.
DataFrame.query - Specify slicing and/or filtering operations dynamically (i.e., as an expression that is evaluated dynamically. Is more applicable to some scenarios than others. Also see this section of the docs for querying on MultiIndexes.
Boolean indexing with a mask generated using MultiIndex.get_level_values (often in conjunction with Index.isin, especially when filtering with multiple values). This is also quite useful in some circumstances.

It will be beneficial to look at the various slicing and filtering problems in terms of the Four Idioms to gain a better understanding what can be applied to a given situation. It is very important to understand that not all of the idioms will work equally well (if at all) in every circumstance. If an idiom has not been listed as a potential solution to a problem below, that means that idiom cannot be applied to that problem effectively.

Question 1
How do I select rows having "a" in level "one"?
         col
one two     
a   t      0
    u      1
    v      2
    w      3

You can use loc, as a general purpose solution applicable to most situations:

df.loc[['a']]

At this point, if you get

TypeError: Expected tuple, got str

That means you're using an older version of pandas. Consider upgrading! Otherwise, use df.loc[('a', slice(None)), :].

Alternatively, you can use xs here, since we are extracting a single cross section. Note the levels and axis arguments (reasonable defaults can be assumed here).

df.xs('a', level=0, axis=0, drop_level=False)
# df.xs('a', drop_level=False)

Here, the drop_level=False argument is needed to prevent xs from dropping level "one" in the result (the level we sliced on).

Yet another option here is using query:

df.query("one == 'a'")

If the index did not have a name, you would need to change your query string to be "ilevel_0 == 'a'".

Finally, using get_level_values:

df[df.index.get_level_values('one') == 'a']
# If your levels are unnamed, or if you need to select by position (not label),
# df[df.index.get_level_values(0) == 'a']

Additionally, how would I be able to drop level "one" in the output?
     col
two     
t      0
u      1
v      2
w      3

This can be easily done using either

df.loc['a'] # Notice the single string argument instead the list.

Or,

df.xs('a', level=0, axis=0, drop_level=True)
# df.xs('a')

Notice that we can omit the drop_level argument (it is assumed to be True by default).

Note

You may notice that a filtered DataFrame may still have all the levels, even if they do not show when printing the DataFrame out. For example,
v = df.loc[['a']]
print(v)
         col
one two     
a   t      0
    u      1
    v      2
    w      3

print(v.index)
MultiIndex(levels=[['a', 'b', 'c', 'd'], ['t', 'u', 'v', 'w']],
           labels=[[0, 0, 0, 0], [0, 1, 2, 3]],
           names=['one', 'two'])
You can get rid of these levels using MultiIndex.remove_unused_levels:
v.index = v.index.remove_unused_levels()

print(v.index)
MultiIndex(levels=[['a'], ['t', 'u', 'v', 'w']],
           labels=[[0, 0, 0, 0], [0, 1, 2, 3]],
           names=['one', 'two'])

Question 1b
How do I slice all rows with value "t" on level "two"?
         col
one two     
a   t      0
b   t      4
    t      8
d   t     12

Intuitively, you would want something involving slice():

df.loc[(slice(None), 't'), :]

It Just Works!™ But it is clunky. We can facilitate a more natural slicing syntax using the pd.IndexSlice API here.

idx = pd.IndexSlice
df.loc[idx[:, 't'], :]

This is much, much cleaner.

Note

Why is the trailing slice : across the columns required? This is because, loc can be used to select and slice along both axes (axis=0 or
axis=1). Without explicitly making it clear which axis the slicing
is to be done on, the operation becomes ambiguous. See the big red box in the documentation on slicing.
If you want to remove any shade of ambiguity, loc accepts an axis
parameter:
df.loc(axis=0)[pd.IndexSlice[:, 't']]
Without the axis parameter (i.e., just by doing df.loc[pd.IndexSlice[:, 't']]), slicing is assumed to be on the columns,
and a KeyError will be raised in this circumstance.
This is documented in slicers. For the purpose of this post, however, we will explicitly specify all axes.

With xs, it is

df.xs('t', axis=0, level=1, drop_level=False)

With query, it is

df.query("two == 't'")
# Or, if the first level has no name, 
# df.query("ilevel_1 == 't'")

And finally, with get_level_values, you may do

df[df.index.get_level_values('two') == 't']
# Or, to perform selection by position/integer,
# df[df.index.get_level_values(1) == 't']

All to the same effect.

Question 2

How can I select rows corresponding to items "b" and "d" in level "one"?

         col
one two     
b   t      4
    u      5
    v      6
    w      7
    t      8
d   w     11
    t     12
    u     13
    v     14
    w     15

Using loc, this is done in a similar fashion by specifying a list.

df.loc[['b', 'd']]

To solve the above problem of selecting "b" and "d", you can also use query:

items = ['b', 'd']
df.query("one in @items")
# df.query("one == @items", parser='pandas')
# df.query("one in ['b', 'd']")
# df.query("one == ['b', 'd']", parser='pandas')

Note

Yes, the default parser is 'pandas', but it is important to highlight this syntax isn't conventionally python. The
Pandas parser generates a slightly different parse tree from the
expression. This is done to make some operations more intuitive to
specify. For more information, please read my post on
Dynamic Expression Evaluation in pandas using pd.eval().

And, with get_level_values + Index.isin:

df[df.index.get_level_values("one").isin(['b', 'd'])]

Question 2b

How would I get all values corresponding to "t" and "w" in level "two"?

         col
one two     
a   t      0
    w      3
b   t      4
    w      7
    t      8
d   w     11
    t     12
    w     15

With loc, this is possible only in conjuction with pd.IndexSlice.

df.loc[pd.IndexSlice[:, ['t', 'w']], :]

The first colon : in pd.IndexSlice[:, ['t', 'w']] means to slice across the first level. As the depth of the level being queried increases, you will need to specify more slices, one per level being sliced across. You will not need to specify more levels beyond the one being sliced, however.

With query, this is

items = ['t', 'w']
df.query("two in @items")
# df.query("two == @items", parser='pandas') 
# df.query("two in ['t', 'w']")
# df.query("two == ['t', 'w']", parser='pandas')

With get_level_values and Index.isin (similar to above):

df[df.index.get_level_values('two').isin(['t', 'w'])]

Question 3
How do I retrieve a cross section, i.e., a single row having a specific values
for the index from df? Specifically, how do I retrieve the cross
section of ('c', 'u'), given by
         col
one two     
c   u      9

Use loc by specifying a tuple of keys:

df.loc[('c', 'u'), :]

Or,

df.loc[pd.IndexSlice[('c', 'u')]]

Note

At this point, you may run into a PerformanceWarning that looks like this:
PerformanceWarning: indexing past lexsort depth may impact performance.
This just means that your index is not sorted. pandas depends on the index being sorted (in this case, lexicographically, since we are dealing with string values) for optimal search and retrieval. A quick fix would be to sort your
DataFrame in advance using DataFrame.sort_index. This is especially desirable from a performance standpoint if you plan on doing
multiple such queries in tandem:
df_sort = df.sort_index()
df_sort.loc[('c', 'u')]
You can also use MultiIndex.is_lexsorted() to check whether the index
is sorted or not. This function returns True or False accordingly.
You can call this function to determine whether an additional sorting
step is required or not.

With xs, this is again simply passing a single tuple as the first argument, with all other arguments set to their appropriate defaults:

df.xs(('c', 'u'))

With query, things become a bit clunky:

df.query("one == 'c' and two == 'u'")

You can see now that this is going to be relatively difficult to generalize. But is still OK for this particular problem.

With accesses spanning multiple levels, get_level_values can still be used, but is not recommended:

m1 = (df.index.get_level_values('one') == 'c')
m2 = (df.index.get_level_values('two') == 'u')
df[m1 & m2]

Question 4
How do I select the two rows corresponding to ('c', 'u'), and ('a', 'w')?
         col
one two     
c   u      9
a   w      3

With loc, this is still as simple as:

df.loc[[('c', 'u'), ('a', 'w')]]
# df.loc[pd.IndexSlice[[('c', 'u'), ('a', 'w')]]]

With query, you will need to dynamically generate a query string by iterating over your cross sections and levels:

cses = [('c', 'u'), ('a', 'w')]
levels = ['one', 'two']
# This is a useful check to make in advance.
assert all(len(levels) == len(cs) for cs in cses) 

query = '(' + ') or ('.join([
    ' and '.join([f"({l} == {repr(c)})" for l, c in zip(levels, cs)]) 
    for cs in cses
]) + ')'

print(query)
# ((one == 'c') and (two == 'u')) or ((one == 'a') and (two == 'w'))

df.query(query)

100% DO NOT RECOMMEND! But it is possible.

What if I have multiple levels?

One option in this scenario would be to use droplevel to drop the levels you're not checking, then use isin to test membership, and then boolean index on the final result.

df[df.index.droplevel(unused_level).isin([('c', 'u'), ('a', 'w')])]

Question 5
How can I retrieve all rows corresponding to "a" in level "one" or
"t" in level "two"?
         col
one two     
a   t      0
    u      1
    v      2
    w      3
b   t      4
    t      8
d   t     12

This is actually very difficult to do with loc while ensuring correctness and still maintaining code clarity. df.loc[pd.IndexSlice['a', 't']] is incorrect, it is interpreted as df.loc[pd.IndexSlice[('a', 't')]] (i.e., selecting a cross section). You may think of a solution with pd.concat to handle each label separately:

pd.concat([
    df.loc[['a'],:], df.loc[pd.IndexSlice[:, 't'],:]
])

         col
one two     
a   t      0
    u      1
    v      2
    w      3
    t      0   # Does this look right to you? No, it isn't!
b   t      4
    t      8
d   t     12

But you'll notice one of the rows is duplicated. This is because that row satisfied both slicing conditions, and so appeared twice. You will instead need to do

v = pd.concat([
        df.loc[['a'],:], df.loc[pd.IndexSlice[:, 't'],:]
])
v[~v.index.duplicated()]

But if your DataFrame inherently contains duplicate indices (that you want), then this will not retain them. Use with extreme caution.

With query, this is stupidly simple:

df.query("one == 'a' or two == 't'")

With get_level_values, this is still simple, but not as elegant:

m1 = (df.index.get_level_values('one') == 'a')
m2 = (df.index.get_level_values('two') == 't')
df[m1 | m2]

Question 6
How can I slice specific cross sections? For "a" and "b", I would like to select all rows with sub-levels "u" and "v", and
for "d", I would like to select rows with sub-level "w".
         col
one two     
a   u      1
    v      2
b   u      5
    v      6
d   w     11
    w     15

This is a special case that I've added to help understand the applicability of the Four Idioms—this is one case where none of them will work effectively, since the slicing is very specific, and does not follow any real pattern.

Usually, slicing problems like this will require explicitly passing a list of keys to loc. One way of doing this is with:

keys = [('a', 'u'), ('a', 'v'), ('b', 'u'), ('b', 'v'), ('d', 'w')]
df.loc[keys, :]

If you want to save some typing, you will recognise that there is a pattern to slicing "a", "b" and its sublevels, so we can separate the slicing task into two portions and concat the result:

pd.concat([
     df.loc[(('a', 'b'), ('u', 'v')), :], 
     df.loc[('d', 'w'), :]
   ], axis=0)

Slicing specification for "a" and "b" is slightly cleaner (('a', 'b'), ('u', 'v')) because the same sub-levels being indexed are the same for each level.

Question 7
How do I get all rows where values in level "two" are greater than 5?
         col
one two     
b   7      4
    9      5
c   7     10
d   6     11
    8     12
    8     13
    6     15

This can be done using query,

df2.query("two > 5")

And get_level_values.

df2[df2.index.get_level_values('two') > 5]

Note

Similar to this example, we can filter based on any arbitrary condition using these constructs. In general, it is useful to remember that loc and xs are specifically for label-based indexing, while query and
get_level_values are helpful for building general conditional masks
for filtering.

Bonus Question
What if I need to slice a MultiIndex column?

Actually, most solutions here are applicable to columns as well, with minor changes. Consider:

np.random.seed(0)
mux3 = pd.MultiIndex.from_product([
        list('ABCD'), list('efgh')
], names=['one','two'])

df3 = pd.DataFrame(np.random.choice(10, (3, len(mux))), columns=mux3)
print(df3)

one  A           B           C           D         
two  e  f  g  h  e  f  g  h  e  f  g  h  e  f  g  h
0    5  0  3  3  7  9  3  5  2  4  7  6  8  8  1  6
1    7  7  8  1  5  9  8  9  4  3  0  3  5  0  2  3
2    8  1  3  3  3  7  0  1  9  9  0  4  7  3  2  7

These are the following changes you will need to make to the Four Idioms to have them working with columns.

To slice with loc, use

 df3.loc[:, ....] # Notice how we slice across the index with `:`.

or,

 df3.loc[:, pd.IndexSlice[...]]

To use xs as appropriate, just pass an argument axis=1.
You can access the column level values directly using df.columns.get_level_values. You will then need to do something like
```
 df.loc[:, {condition}] 
```
Where {condition} represents some condition built using columns.get_level_values.
To use query, your only option is to transpose, query on the index, and transpose again:
```
 df3.T.query(...).T
```
Not recommended, use one of the other 3 options.

The right way to query a pandas MultiIndex

I think what you did is fine, but there are alternative ways also.

>>> df = pd.DataFrame({ 
              'stock':np.repeat( ['AAPL','GOOG','YHOO'], 3 ),
              'date':np.tile( pd.date_range('5/5/2015', periods=3, freq='D'), 3 ),
              'price':(np.random.randn(9).cumsum() + 10) })

>>> df = df.set_index(['stock','date'])

                      price
stock date                 
AAPL  2015-05-05   8.538459
      2015-05-06   9.330140
      2015-05-07   8.968898
GOOG  2015-05-05   8.964389
      2015-05-06   9.828230
      2015-05-07   9.992985
YHOO  2015-05-05   9.929548
      2015-05-06   9.330295
      2015-05-07  10.676468

A slightly more standard way of using loc twice

>>> df.loc['AAPL'].loc['2015-05-05']

would be to do

>>> df.loc['AAPL','2015-05-05']

price    8.538459
Name: (AAPL, 2015-05-05 00:00:00), dtype: float64

And instead of xs you could use an IndexSlice. I think for 2 levels xs is easier, but IndexSlice might be better past 2 levels.

>>> idx=pd.IndexSlice

>>> df.loc[ idx[:,'2015-05-05'], : ]

                     price
stock date                
AAPL  2015-05-05  8.538459
GOOG  2015-05-05  8.964389
YHOO  2015-05-05  9.929548

And to be honest, I think the absolute easiest way here is use either date or stock (or neither) as index and then most selections are very straightforward. For example, if you remove the index completely you can effortlessly select by date:

>>> df = df.reset_index()
>>> df[ df['date']=='2015-05-05' ]

   index stock       date      price
0      0  AAPL 2015-05-05   8.538459
3      3  GOOG 2015-05-05   8.964389
6      6  YHOO 2015-05-05   9.929548

Doing some quickie timings with 3 stocks and 3000 dates (=9000 rows), I found that a simple boolean selection (no index) was about 35% faster than xs, and xs was about 35% faster than using IndexSlice. But see Jeff's comment below, you should expect the boolean selection to perform relative worse with more rows.

Of course, the best thing for you to do is test on your own data and see how it comes out.

selecting from multi-index pandas

One way is to use the get_level_values Index method:

In [11]: df
Out[11]:
     0
A B
1 4  1
2 5  2
3 6  3

In [12]: df.iloc[df.index.get_level_values('A') == 1]
Out[12]:
     0
A B
1 4  1

In 0.13 you'll be able to use xs with drop_level argument:

df.xs(1, level='A', drop_level=False) # axis=1 if columns

Note: if this were column MultiIndex rather than index, you could use the same technique:

In [21]: df1 = df.T

In [22]: df1.iloc[:, df1.columns.get_level_values('A') == 1]
Out[22]:
A  1
B  4
0  1

pandas dataframe select columns in multiindex

There is a get_level_values method that you can use in conjunction with boolean indexing to get the the intended result.

In [13]:

df = pd.DataFrame(np.random.random((4,4)))
df.columns = pd.MultiIndex.from_product([[1,2],['A','B']])
print df
          1                   2          
          A         B         A         B
0  0.543980  0.628078  0.756941  0.698824
1  0.633005  0.089604  0.198510  0.783556
2  0.662391  0.541182  0.544060  0.059381
3  0.841242  0.634603  0.815334  0.848120
In [14]:

print df.iloc[:, df.columns.get_level_values(1)=='A']
          1         2
          A         A
0  0.543980  0.756941
1  0.633005  0.198510
2  0.662391  0.544060
3  0.841242  0.815334

Querying MultiIndex DataFrame in Pandas

Use merge with parameter left_index and right_on:

df = FirstDF.merge(SecondDF, left_index=True, right_on=['A','B'])['C'].to_frame()
print (df)
    C
0  59
1  56
2  80

Another solution with isin of MultiIndexes and filtering by boolean indexing:

mask = FirstDF.index.isin(SecondDF.set_index(['A','B']).index)
#alternative solution
#mask = FirstDF.index.isin(list(map(tuple,SecondDF[['A','B']].values.tolist())))
df = FirstDF.loc[mask, ['C']].reset_index(drop=True)
print (df)
    C
0  59
1  56
2  80

Detail:

print (FirstDF.loc[mask, ['C']])
              C
A   B          
'a' 'green'  59
'b' 'red'    56
'c' 'green'  80

EDIT:

You can use merge with outer join and indicator=True parameter, then filter by boolean indexing:

df1=FirstDF.merge(SecondDF, left_index=True, right_on=['A','B'], indicator=True, how='outer')
print (df1)
    C    A         B     _merge
2  43  'a'    'blue'  left_only
0  59  'a'   'green'       both
1  56  'b'     'red'       both
2  80  'c'   'green'       both
2  72  'c'  'orange'  left_only

mask = df1['_merge'] != 'both'
df1 = df1.loc[mask, ['C']].reset_index(drop=True)
print (df1)
    C
0  43
1  72

For second solution invert boolen mask by ~:

mask = FirstDF.index.isin(SecondDF.set_index(['A','B']).index)
#alternative solution
#mask = FirstDF.index.isin(list(map(tuple,SecondDF[['A','B']].values.tolist())))
df = FirstDF.loc[~mask, ['C']].reset_index(drop=True)
print (df)
    C
0  43
1  72

Pandas multiindex: select by condition

Use tuples in DataFrame.loc:

a = df.loc[df[("cat2", "vals2")] == 7, ('cat1', 'vals1')]
print (a)

Last if need scalar from one element Series:

out = a.iat[0]

If possible no match:

out = next(iter(a), 'no match')

You can compare sliced rows, columns - output is DataFrame filled by boolean - so for boolean Series need test if any True per row by DataFrame.any or all Trues per rows by DataFrame.all:

m = df.loc[:, ("cat2", slice(None))]==7

a = df.loc[m.any(axis=1), ("cat1", "vals2")]
print (a)
a  d    5
Name: (cat1, vals2), dtype: int32

m = df.loc[:, ("cat2", slice(None))]==7
df2 = df.loc[m.any(axis=1), ("cat1", slice(None))]
print (df2)
     cat1      
    vals1 vals2
a d     4     5

How to Query Multiindex Index Columns Values in Pandas