python pandas filter function regex behavior on MultiIndex dataframe

Question

I have a dataframe df that looks like (see Appendix for code to generate the dataframe):

fy 2018 2019 tag uom Assets USD 3.753190e+11 3.385160e+11 AssetsCurrent USD 1.286450e+11 1.628190e+11 AssetsNoncurrent USD 2.466740e+11 1.756970e+11 DeferredTaxAssetsDeferredCostSharing USD 6.670000e+08 NaN DeferredTaxAssetsDeferredIncome USD 1.521000e+09 1.141000e+09 DeferredTaxAssetsGoodwillAndIntangibleAssets USD NaN 1.143300e+10 DeferredTaxAssetsLiabilitiesNet USD 5.834000e+09 5.834000e+09 DeferredTaxAssetsNet USD 8.974000e+09 6.610000e+09 DeferredTaxAssetsOther USD 8.340000e+08 7.970000e+08 DeferredTaxAssetsPropertyPlantAndEquipment USD 1.230000e+09 1.370000e+08 DeferredTaxAssetsTaxDeferredExpenseCompensation... USD 7.030000e+08 5.130000e+08 DeferredTaxAssetsTaxDeferredExpenseReservesAndA... USD 4.019000e+09 3.151000e+09 DeferredTaxAssetsUnrealizedLossesOnAvailablefor... USD 0.000000e+00 8.710000e+08 DerivativeAssetsReductionforMasterNettingArrang... USD 1.400000e+09 2.100000e+09 IncreaseDecreaseInOtherOperatingAssets USD -1.055000e+09 5.318000e+09 NoncurrentAssets USD 3.378300e+10 4.130400e+10 OtherAssetsCurrent USD 1.208700e+10 1.208700e+10 OtherAssetsNoncurrent USD 2.228300e+10 2.228300e+10

Which is a MultiIndex pivot table with indices tag and uom . My goal is to filter rows by the tag index using a regex and the filter function . For example:

df.filter(regex="^Assets$", axis="index")

Which ideally would filter out the row:

fy 2018 2019 tag uom Assets USD 3.753190e+11 3.385160e+11

However, when I do so it outputs an empty dataframe:

Empty DataFrame Columns: [2018, 2019] Index: []

I'm able to circumvent this problem by using:

df.index.get_level_values("tag").str.contains("^Assets$")

or as a function

search = lambda df, regex, index_name: df.loc[df.index.get_level_values(index_name).str.contains(regex)]

But this is way less satisfying to me. Am I missing something about the pandas filter function and how its regex input works? It does not behave as expected, and my guess is it's because I have 2 indices: tag and uom thus the regex is failing in the uom index when I use "^Assets$" as my regex. This is supported by using the regex "^Assets$|USD" which returns the entire dataframe because all rows have uom=USD , and it shows the filter function takes both indices into account. If this is the case, then how do I selectively choose index= tag for the filter function on a MultiIndex dataframe?

Appendix:

import pandas as pd
import numpy as np

levels = ['Assets',
         'AssetsCurrent',
         'AssetsNoncurrent',
         'DeferredTaxAssetsDeferredCostSharing',
         'DeferredTaxAssetsDeferredIncome',
         'DeferredTaxAssetsGoodwillAndIntangibleAssets',
         'DeferredTaxAssetsLiabilitiesNet',
         'DeferredTaxAssetsNet',
         'DeferredTaxAssetsOther',
         'DeferredTaxAssetsPropertyPlantAndEquipment',
'DeferredTaxAssetsTaxDeferredExpenseCompensationAndBenefitsShareBasedCompensationCost',
         'DeferredTaxAssetsTaxDeferredExpenseReservesAndAccruals',
         'DeferredTaxAssetsUnrealizedLossesOnAvailableforSaleSecuritiesGross',
         'DerivativeAssetsReductionforMasterNettingArrangements',
         'IncreaseDecreaseInOtherOperatingAssets',
         'NoncurrentAssets',
         'OtherAssetsCurrent',
         'OtherAssetsNoncurrent']
codes = ['USD' for i in range(len(levels))]

index = pd.MultiIndex.from_arrays([levels, codes], names=['tag', 'uom'])
columns = pd.Int64Index([2018, 2019], dtype='int64', name='fy')
values = [[3.75319e+11,  3.38516e+11],
              [1.28645e+11, 1.62819e+11],
              [2.46674e+11,  1.75697e+11],
              [6.67000e+08,          np.NaN],
              [1.52100e+09,  1.14100e+09],
              [np.NaN,  1.14330e+10],
              [5.83400e+09,  5.83400e+09],
              [8.97400e+09,  6.61000e+09],
              [8.34000e+08,  7.97000e+08],
              [1.23000e+09,  1.37000e+08],
              [7.03000e+08,  5.13000e+08],
              [4.01900e+09,  3.15100e+09],
              [0.00000e+00,  8.71000e+08],
              [1.40000e+09,  2.10000e+09],
              [-1.05500e+09,  5.31800e+09],
              [3.37830e+10,  4.13040e+10],
              [1.20870e+10,  1.20870e+10],
              [2.22830e+10,  2.22830e+10]]

df = pd.DataFrame(values, columns=columns, index=index)

Answer 1

The implementation of the regex part of the filter function is short and easy to adapt for a multi-index scenario where you still want to only regex 1 part of the multi-index. I know this is not a direct answer to what you asked because you're right, as implemented the filter function does not handle multi-index.

I ended up here with the same problem and thought it might be a useful answer to others to post the code I have used, adapted from the pandas original:

import regex as re

def filter_multi(df, index_level_name, regex, axis=0):
    def f(x):
        return matcher.search(str(x)) is not None

    matcher = re.compile(regex)
    values = df.axes[axis].get_level_values(index_level_name).map(f)
    return df.loc(axis=axis)[values]

Using the code in your Appendix:

print(df)
print(filter_multi(df, index_level_name='tag', regex='^Assets$', axis=0))
print(filter_multi(df, index_level_name='fy', regex='^2019$', axis=1))

Answer 2

If you want to filter a unique value from the first part of a multi-index, you can just use loc :

df.loc[['Assets']]

which gives:

fy                  2018          2019
tag    uom                            
Assets USD  3.753190e+11  3.385160e+11

If for your real problem, filter must be used, you should reset the unused part of the index and set it back after filtering:

df.reset_index(level='uom').filter(regex='^Assets$', axis=0).set_index('uom', append=True)

Answer 3

Another option is to first remove uom from your index, apply filter (which then will be applied to the only index tag ) and add uom back to your index, as in

df.reset_index('uom').filter(regex="^Assets$", axis=0).set_index('uom', append=True)