Pandas MultiIndex 重新排列列

Question

The MultiIndex rearranges the columns seemingly randomly when the label values are not aligned, when I use the function get_level_values to get the columns values.

For instance, I can create a MultiIndex, whose labels are ordered from 0 to 4.

import pandas as pd
import numpy as np
work_index=pd.MultiIndex(levels=[['ANA','ANC','PPI','SCAF','SAC'],['Sample']],labels=[[0,1,2,3,4],[0,0,0,0,0]])

When I put this MultiIndex into a dataframe and run the get_levels_values function I get ['ANA','ANC','PPI','SCAF','SAC'] in the proper order I expect to get.

work=pd.DataFrame(np.random.randn(5,5),columns=work_index)
work.columns #note the proper order
>>> MultiIndex(levels=[['ANA', 'ANC', 'PPI', 'SCAF', 'SAC'], ['Sample']],
           labels=[[0, 1, 2, 3, 4], [0, 0, 0, 0, 0]])
work.columns.get_level_values(0) #same order as before
>>> Index(['ANA', 'ANC', 'PPI', 'SCAF', 'SAC'], dtype='object')

However, if I create a multindex with the labels not in numerical order, get_level_values returns a weird, seemingly random order. Here instead of [0,1,2,3,4] I choose [2,1,4,3,0] .

not_work_index=pd.MultiIndex(levels=[['ANA','ANC','PPI','SCAF','SAC'],['Sample']],labels=[[2, 1, 4, 3, 0],[0,0,0,0,0]])

Putting this into a dataframe does not give me ['ANA','ANC','PPI','SCAF','SAC'] , rather ['PPI','ANC','SAC','SCAF','ANA']

not_work=pd.DataFrame(np.random.randn(5,5),columns=not_work_index)
not_work.columns
>>> MultiIndex(levels=[['ANA', 'ANC', 'PPI', 'SCAF', 'SAC'], ['Sample']],
       labels=[[2, 1, 4, 3, 0], [0, 0, 0, 0, 0]])
not_work.columns.get_level_values(0)
>>> Index(['PPI', 'ANC', 'SAC', 'SCAF', 'ANA'], dtype='object')

Is there a way for get_level_values to return the levels in order even if the labels are not in order? Is there another way to query the upper level to get the columns in correct order?

Answer 1

I'm not sure if this is a bug or not, it looks like get_level_values always returns a sorted array ignoring the creation order, the IndexArray itself knows the correct order. You can get the order you want using the following gnarly code to get the label array to mask the level values:

In [11]:
not_work.columns.get_level_values(0)[not_work.columns.labels[0]]

Out[11]:
Index(['SAC', 'ANC', 'ANA', 'SCAF', 'PPI'], dtype='object')

Here I access the labels attribute of the IndexArray or columns:

In [12]:
not_work.columns.labels

Out[12]:
FrozenList([[2, 1, 4, 3, 0], [0, 0, 0, 0, 0]])

I then index the first level using [0] :

In [13]:
not_work.columns.labels[0]

Out[13]:
FrozenNDArray([2, 1, 4, 3, 0], dtype='int8')

We can then use this to mask the level values to return the original order:

In [11]:
not_work.columns.get_level_values(0)[not_work.columns.labels[0]]

Out[11]:
Index(['SAC', 'ANC', 'ANA', 'SCAF', 'PPI'], dtype='object')