5

I have a dataframe with several string columns that I want to convert to categorical data so that I can run some models and extract important features from.

However, due to the amount of unique values, the one-hot encoded data expands into a large number of columns which is causing performance issues.

To combat this, I'm experimenting with the Sparse = True parameter in get_dummies. 

test1 = pd.get_dummies(X.loc[:,['col1','col2','col3','col4']].head(10000))
test2 = pd.get_dummies(X.loc[:,['col1','col2','col3','col4']].head(10000),sparse = True)

However, when I check info for my two comparison objects, they take up the same amount of memory. It doesn't seem like Sparse = True uses any less space. Why is that?

test1.info()

<class 'pandas.core.frame.DataFrame'>
Int64Index: 10000 entries, 537293 to 752152
Columns: 2253 entries,...
dtypes: uint8(2253)
memory usage: 21.6 MB

test2.info()
<class 'pandas.core.sparse.frame.SparseDataFrame'>
Int64Index: 10000 entries, 537293 to 752152
Columns: 2253 entries, ...
dtypes: uint8(2253)
memory usage: 21.9 MB
trystuff
  • 686
  • 1
  • 8
  • 18

1 Answers1

6

I looked at pandas get_dummies source but could not spot an error so far. Here is a small experiment that I did below (1st half is reproducing your problem with real data).

In [1]: import numpy as np
   ...: import pandas as pd
   ...: 
   ...: a = ['a', 'b'] * 100000
   ...: A = ['A', 'B'] * 100000
   ...: 
   ...: df1 = pd.DataFrame({'a': a, 'A': A})
   ...: df1 = pd.get_dummies(df1)
   ...: df1.info()
   ...:
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 200000 entries, 0 to 199999
Data columns (total 4 columns):
A_A    200000 non-null uint8
A_B    200000 non-null uint8
a_a    200000 non-null uint8
a_b    200000 non-null uint8
dtypes: uint8(4)
memory usage: 781.3 KB

In [2]: df2 = pd.DataFrame({'a': a, 'A': A})
   ...: df2 = pd.get_dummies(df2, sparse=True)
   ...: df2.info()
   ...:
<class 'pandas.core.sparse.frame.SparseDataFrame'>
RangeIndex: 200000 entries, 0 to 199999
Data columns (total 4 columns):
A_A    200000 non-null uint8
A_B    200000 non-null uint8
a_a    200000 non-null uint8
a_b    200000 non-null uint8
dtypes: uint8(4)
memory usage: 781.3 KB

So far the same result (the size of df1 is equal to that of df2) as yours, but if I explicitly convert df2 to sparse using to_sparse with fill_value=0

In [3]: df2 = df2.to_sparse(fill_value=0)
   ...: df2.info()
   ...:
<class 'pandas.core.sparse.frame.SparseDataFrame'>
RangeIndex: 200000 entries, 0 to 199999
Data columns (total 4 columns):
A_A    200000 non-null uint8
A_B    200000 non-null uint8
a_a    200000 non-null uint8
a_b    200000 non-null uint8
dtypes: uint8(4)
memory usage: 390.7 KB

Now the memory usage is half since half of the data is 0.

In conclusion, I'm not sure why get_dummies(sparse=True) does not compress the dataframe even though it is converted to SparseDataFrame, but there is a workaround. Related discussion was going on in github get_dummies with sparse doesn't convert numeric to sparse but the conclusion still seems to be up in the air.

gyoza
  • 2,112
  • 2
  • 12
  • 18