Search code examples
pythonpandasmulti-index

Using boolean indexing for row and column MultiIndex in Pandas

Questions are at the end, in bold. But first, let's set up some data:

import numpy as np
import pandas as pd
from itertools import product

np.random.seed(1)

team_names = ['Yankees', 'Mets', 'Dodgers']
jersey_numbers = [35, 71, 84]
game_numbers = [1, 2]
observer_names = ['Bill', 'John', 'Ralph']
observation_types = ['Speed', 'Strength']

row_indices = list(product(team_names, jersey_numbers, game_numbers, observer_names, observation_types))
observation_values = np.random.randn(len(row_indices))

tns, jns, gns, ons, ots = zip(*row_indices)

data = pd.DataFrame({'team': tns, 'jersey': jns, 'game': gns, 'observer': ons, 'obstype': ots, 'value': observation_values})

data = data.set_index(['team', 'jersey', 'game', 'observer', 'obstype'])
data = data.unstack(['observer', 'obstype'])
data.columns = data.columns.droplevel(0)

this gives: data

I want to pluck out a subset of this DataFrame for subsequent analysis. Say I wanted to slice out the rows where the jersey number is 71. I don't really like the idea of using xs to do this. When you do a cross section via xs you lose the column you selected on. If I run:

data.xs(71, axis=0, level='jersey')

then I get back the right rows, but I lose the jersey column.

xs_slice

Also, xs doesn't seem like a great solution for the case where I want a few different values from the jersey column. I think a much nicer solution is the one found here:

data[[j in [71, 84] for t, j, g in data.index]]

boolean_slice_1

You could even filter on a combination of jerseys and teams:

data[[j in [71, 84] and t in ['Dodgers', 'Mets'] for t, j, g in data.index]]

boolean_slice_2

Nice!

So the question: how can I do something similar for selecting a subset of columns. For example, say I want only the columns representing data from Ralph. How can I do that without using xs? Or what if I wanted only the columns with observer in ['John', 'Ralph']? Again, I'd really prefer a solution that keeps all the levels of the row and column indices in the result...just like the boolean indexing examples above.

I can do what I want, and even combine selections from both the row and column indices. But the only solution I've found involves some real gymnastics:

data[[j in [71, 84] and t in ['Dodgers', 'Mets'] for t, j, g in data.index]]\
    .T[[obs in ['John', 'Ralph'] for obs, obstype in data.columns]].T

double_boolean_slice

And thus the second question: is there a more compact way to do what I just did above?

Solution

  • Here is one approach that uses slightly more built-in-feeling syntax. But it's still clunky as hell:

    data.loc[
    (data.index.get_level_values('jersey').isin([71, 84])
     & data.index.get_level_values('team').isin(['Dodgers', 'Mets'])), 
    data.columns.get_level_values('observer').isin(['John', 'Ralph'])
]

    So comparing:

    def hackedsyntax():
    return data[[j in [71, 84] and t in ['Dodgers', 'Mets'] for t, j, g in data.index]]\
    .T[[obs in ['John', 'Ralph'] for obs, obstype in data.columns]].T

def uglybuiltinsyntax():
    return data.loc[
        (data.index.get_level_values('jersey').isin([71, 84])
         & data.index.get_level_values('team').isin(['Dodgers', 'Mets'])), 
        data.columns.get_level_values('observer').isin(['John', 'Ralph'])
    ]

%timeit hackedsyntax()
%timeit uglybuiltinsyntax()

hackedsyntax() - uglybuiltinsyntax()

    results:

    1000 loops, best of 3: 395 µs per loop
1000 loops, best of 3: 409 µs per loop

    comparison_of_methods

    Still hopeful there's a cleaner or more canonical way to do this.