pandas - pivot_table with non-numeric values? (DataError: No numeric types to aggregate)

Question

I'm trying to do a pivot of a table containing strings as results.

import pandas as pd

df1 = pd.DataFrame({'index' : range(8),
'variable1' : ["A","A","B","B","A","B","B","A"],
'variable2' : ["a","b","a","b","a","b","a","b"],
'variable3' : ["x","x","x","y","y","y","x","y"],
'result': ["on","off","off","on","on","off","off","on"]})

df1.pivot_table(values='result',rows='index',cols=['variable1','variable2','variable3'])

But I get: DataError: No numeric types to aggregate.

This works as intended when I change result values to numbers:

df2 = pd.DataFrame({'index' : range(8),
'variable1' : ["A","A","B","B","A","B","B","A"],
'variable2' : ["a","b","a","b","a","b","a","b"],
'variable3' : ["x","x","x","y","y","y","x","y"],
'result': [1,0,0,1,1,0,0,1]})

df2.pivot_table(values='result',rows='index',cols=['variable1','variable2','variable3'])

And I get what I need:

variable1   A               B    
variable2   a       b       a   b
variable3   x   y   x   y   x   y
index                            
0           1 NaN NaN NaN NaN NaN
1         NaN NaN   0 NaN NaN NaN
2         NaN NaN NaN NaN   0 NaN
3         NaN NaN NaN NaN NaN   1
4         NaN   1 NaN NaN NaN NaN
5         NaN NaN NaN NaN NaN   0
6         NaN NaN NaN NaN   0 NaN
7         NaN NaN NaN   1 NaN NaN

I know I can map the strings to numerical values and then reverse the operation, but maybe there is a more elegant solution?

Answer 1

My original reply was based on Pandas 0.14.1, and since then, many things changed in the pivot_table function (rows --> index, cols --> columns... )

Additionally, it appears that the original lambda trick I posted no longer works on Pandas 0.18. You have to provide a reducing function (even if it is min, max or mean). But even that seemed improper - because we are not reducing the data set, just transforming it.... So I looked harder at unstack...

import pandas as pd

df1 = pd.DataFrame({'index' : range(8),
'variable1' : ["A","A","B","B","A","B","B","A"],
'variable2' : ["a","b","a","b","a","b","a","b"],
'variable3' : ["x","x","x","y","y","y","x","y"],
'result': ["on","off","off","on","on","off","off","on"]})

# these are the columns to end up in the multi-index columns.
unstack_cols = ['variable1', 'variable2', 'variable3']

First, set an index on the data using the index + the columns you want to stack, then call unstack using the level arg.

df1.set_index(['index'] + unstack_cols).unstack(level=unstack_cols)

Resulting dataframe is below.

Answer 2

I think the best compromise is to replace on/off with True/False, which will enable pandas to "understand" the data better and act in an intelligent, expected way.

df2 = df1.replace({'on': True, 'off': False})

You essentially conceded this in your question. My answer is, I don't think there's a better way, and you should replace 'on'/'off' anyway for whatever comes next.

As Andy Hayden points out in the comments, you'll get better performance if you replace on/off with 1/0.