pandas dataframe columns scaling with sklearn

Question

I have a pandas dataframe with mixed type columns, and I'd like to apply sklearn's min_max_scaler to some of the columns. Ideally, I'd like to do these transformations in place, but haven't figured out a way to do that yet. I've written the following code that works:

import pandas as pd
import numpy as np
from sklearn import preprocessing

scaler = preprocessing.MinMaxScaler()

dfTest = pd.DataFrame({'A':[14.00,90.20,90.95,96.27,91.21],'B':[103.02,107.26,110.35,114.23,114.68], 'C':['big','small','big','small','small']})
min_max_scaler = preprocessing.MinMaxScaler()

def scaleColumns(df, cols_to_scale):
    for col in cols_to_scale:
        df[col] = pd.DataFrame(min_max_scaler.fit_transform(pd.DataFrame(dfTest[col])),columns=[col])
    return df

dfTest

    A   B   C
0    14.00   103.02  big
1    90.20   107.26  small
2    90.95   110.35  big
3    96.27   114.23  small
4    91.21   114.68  small

scaled_df = scaleColumns(dfTest,['A','B'])
scaled_df

A   B   C
0    0.000000    0.000000    big
1    0.926219    0.363636    small
2    0.935335    0.628645    big
3    1.000000    0.961407    small
4    0.938495    1.000000    small

I'm curious if this is the preferred/most efficient way to do this transformation. Is there a way I could use df.apply that would be better?

I'm also surprised I can't get the following code to work:

bad_output = min_max_scaler.fit_transform(dfTest['A'])

If I pass an entire dataframe to the scaler it works:

dfTest2 = dfTest.drop('C', axis = 1)
good_output = min_max_scaler.fit_transform(dfTest2)
good_output

I'm confused why passing a series to the scaler fails. In my full working code above I had hoped to just pass a series to the scaler then set the dataframe column = to the scaled series.

Solution 1

I am not sure if previous versions of pandas prevented this but now the following snippet works perfectly for me and produces exactly what you want without having to use apply

>>> import pandas as pd
>>> from sklearn.preprocessing import MinMaxScaler


>>> scaler = MinMaxScaler()

>>> dfTest = pd.DataFrame({'A':[14.00,90.20,90.95,96.27,91.21],
                           'B':[103.02,107.26,110.35,114.23,114.68],
                           'C':['big','small','big','small','small']})

>>> dfTest[['A', 'B']] = scaler.fit_transform(dfTest[['A', 'B']])

>>> dfTest
          A         B      C
0  0.000000  0.000000    big
1  0.926219  0.363636  small
2  0.935335  0.628645    big
3  1.000000  0.961407  small
4  0.938495  1.000000  small

Solution 2

Like this?

dfTest = pd.DataFrame({
           'A':[14.00,90.20,90.95,96.27,91.21],
           'B':[103.02,107.26,110.35,114.23,114.68], 
           'C':['big','small','big','small','small']
         })
dfTest[['A','B']] = dfTest[['A','B']].apply(
                           lambda x: MinMaxScaler().fit_transform(x))
dfTest

    A           B           C
0   0.000000    0.000000    big
1   0.926219    0.363636    small
2   0.935335    0.628645    big
3   1.000000    0.961407    small
4   0.938495    1.000000    small

Solution 3

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

This should work without depreciation warnings.

Solution 4

As it is being mentioned in pir's comment - the .apply(lambda el: scale.fit_transform(el)) method will produce the following warning:

DeprecationWarning: Passing 1d arrays as data is deprecated in 0.17 and will raise ValueError in 0.19. Reshape your data either using X.reshape(-1, 1) if your data has a single feature or X.reshape(1, -1) if it contains a single sample.

Converting your columns to numpy arrays should do the job (I prefer StandardScaler):

from sklearn.preprocessing import StandardScaler
scale = StandardScaler()

dfTest[['A','B','C']] = scale.fit_transform(dfTest[['A','B','C']].as_matrix())

-- Edit Nov 2018 (Tested for pandas 0.23.4)--

As Rob Murray mentions in the comments, in the current (v0.23.4) version of pandas .as_matrix() returns FutureWarning. Therefore, it should be replaced by .values:

from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()

scaler.fit_transform(dfTest[['A','B']].values)

-- Edit May 2019 (Tested for pandas 0.24.2)--

As joelostblom mentions in the comments, "Since 0.24.0, it is recommended to use .to_numpy() instead of .values."

Updated example:

import pandas as pd
from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()
dfTest = pd.DataFrame({
               'A':[14.00,90.20,90.95,96.27,91.21],
               'B':[103.02,107.26,110.35,114.23,114.68],
               'C':['big','small','big','small','small']
             })
dfTest[['A', 'B']] = scaler.fit_transform(dfTest[['A','B']].to_numpy())
dfTest
      A         B      C
0 -1.995290 -1.571117    big
1  0.436356 -0.603995  small
2  0.460289  0.100818    big
3  0.630058  0.985826  small
4  0.468586  1.088469  small

Solution 5

You can do it using pandas only:

In [235]:
dfTest = pd.DataFrame({'A':[14.00,90.20,90.95,96.27,91.21],'B':[103.02,107.26,110.35,114.23,114.68], 'C':['big','small','big','small','small']})
df = dfTest[['A', 'B']]
df_norm = (df - df.min()) / (df.max() - df.min())
print df_norm
print pd.concat((df_norm, dfTest.C),1)

          A         B
0  0.000000  0.000000
1  0.926219  0.363636
2  0.935335  0.628645
3  1.000000  0.961407
4  0.938495  1.000000
          A         B      C
0  0.000000  0.000000    big
1  0.926219  0.363636  small
2  0.935335  0.628645    big
3  1.000000  0.961407  small
4  0.938495  1.000000  small

View more solutions

Comments

[EdChum]

Does it work if you do this bad_output = min_max_scaler.fit_transform(dfTest['A'].values)? accessing the values attribute returns a numpy array, for some reason sometimes the scikit learn api will correctly call the right method that makes pandas returns a numpy array and sometimes it doesn't.

[Fred Foo]

Pandas' dataframes are quite complicated objects with conventions that do not match scikit-learn's conventions. If you convert everything to NumPy arrays, scikit-learn gets a lot easier to work with.

[flyingmeatball]

@edChum - bad_output = in_max_scaler.fit_transform(dfTest['A'].values) did not work either. @larsmans - yeah I had thought about going down this route, it just seems like a hassle. I don't know if it is a bug or not that Pandas can pass a full dataframe to a sklearn function, but not a series. My understanding of a dataframe was that it is a dict of series. Reading in the "Python for Data Analysis" book, it states that pandas is built on top of numpy to make it easy to use in NumPy-centric applicatations.

[flyingmeatball]

I know that I can do it just in pandas, but I may want to eventually apply a different sklearn method that isn't as easy to write myself. I'm more interested in figuring out why applying to a series doesn't work as I expected than I am in coming up with a strictly simpler solution. My next step will be to run a RandomForestRegressor, and I want to make sure I understand how Pandas and sklearn work together.

[pir]

I get a bunch of DeprecationWarnings when I run this script. How should it be updated?

[AJP]

See @LetsPlayYahtzee's answer below

[citynorman]

Neat! A more generalized version df[df.columns] = scaler.fit_transform(df[df.columns])

[Rajesh Mappu]

I know this is a delayed comment from original date, but why is there two square brackets in dfTest[['A', 'B']]? I can see it doesn't work with single bracket, but couldn't understand the reason.

[ken]

@RajeshThevar The outer brackets are pandas' typical selector brackets, telling pandas to select a column from the dataframe. The inner brackets indicate a list. You're passing a list to the pandas selector. If you just use single brackets - with one column name followed by another, separated by a comma - pandas interprets this as if you're trying to select a column from a dataframe with multi-level columns (a MultiIndex) and will throw a keyerror.

[LetsPlayYahtzee]

to add to @ken's answer if you want to see exactly how pandas implements this indexing logic and why a tuple of values would be interpreted differently than a list you can look at how DataFrames implement the __getitem__ method. Specifically you can open you ipython and do pd.DataFrame.__getitem__?? ; after you import pandas as pd of course ;)

[Adam Stewart]

This works great, however when I try to do scaler.fit() and scaler.transform() on two separate lines I'm getting the dreaded SettingWithCopyWarning. Anyone have any idea why?

[David J.]

A practical note: for those using train/test data splits, you'll want to only fit on your training data, not your testing data.

[satsumas]

df[df.columns] = scaler.fit_transform(df[df.columns]) -- perfect @citynorman!

[Asclepius]

This answer is dangerous because df.max() - df.min() can be 0, leading to an exception. Moreover, df.min() is computed twice which is inefficient. Note that df.ptp() is equivalent to df.max() - df.min().

[Alexandre V.]

A simpler version: dfTest[['A','B']] = dfTest[['A','B']].apply(MinMaxScaler().fit_transform)

[Rob Murray]

use .values in place of .as_matrix() as as_matrix() now gives a FutureWarning.

[intotecho]

To scale all but the timestamps column, combine with columns =df.columns.drop('timestamps') df[df.columns] = scaler.fit_transform(df[df.columns]

[compguy24]

@DavidJ. can you provide a reference for why?

[joelostblom]

Since 0.24.0, it is recommended to use .to_numpy() instead of .values.

[JolonB]

Correction of @intotecho's comment. You should do columns = df.columns.drop('timestamps') and df[columns] = scaler.fit_transform(df[columns]). It should be columns in the square brackets, not df.columns

[HFX]

How to inverse a value in this way?

[shcrela]

or df[df.columns] = scale.fit_transform(df)

[Ammanuel]

Works perfectly! I was trying to figure out how to retain the column names, this helped.

[ssp]

Can use df[:] = scaler.fit_transform(df) if scaling the entire dataset.

[ICW]

this will instantiate a new MinMaxScaler per row not sure if it matters though