whitespace is a terrible separator for feature names in PolynomialFeatures.get_feature_names

[amueller]

If your feature names have white spaces in them, it's hard to see which features are interactions right now.
Maybe we should make the separator an option of get_feature_names or use something like *?

[mohamed-ali]

@amueller I will try to reproduce the issue, if confirmed I will push a PR to fix it by adding separator as an option.

[mohamed-ali]

Here's an example that reproduces the issue

>>> import pandas as pd 
>>> df = pd.DataFrame([[1, 2, 3], [4, 5, 6], [7, 8, 9]], columns=["A B", "C", "D"])
>>> df
   A B  C  D
0    1  2  3
1    4  5  6
2    7  8  9
>>> from sklearn.preprocessing import PolynomialFeatures
>>> poly = PolynomialFeatures(2)
>>> poly.fit(df)
PolynomialFeatures(degree=2, include_bias=True, interaction_only=False)
>>> poly.transform(df)
array([[  1.,   1.,   2.,   3.,   1.,   2.,   3.,   4.,   6.,   9.],
       [  1.,   4.,   5.,   6.,  16.,  20.,  24.,  25.,  30.,  36.],
       [  1.,   7.,   8.,   9.,  49.,  56.,  63.,  64.,  72.,  81.]])
>>> poly.get_feature_names() 
['1', 'x0', 'x1', 'x2', 'x0^2', 'x0 x1', 'x0 x2', 'x1^2', 'x1 x2', 'x2^2']
>>> poly.get_feature_names(input_features=df.columns.tolist()) 
['1', 'A B', 'C', 'D', 'A B^2', 'A B C', 'A B D', 'C^2', 'C D', 'D^2']

When including a feature which has a whitespace, in this case the feature "A B" output is ambiguous.

[mohamed-ali]

Changing the code in here from

name = " ".join("%s^%d" % (input_features[ind], exp)
                                if exp != 1 else input_features[ind]
                                for ind, exp in zip(inds, row[inds]))

to this

name = "*".join("%s^%d" % (input_features[ind], exp)
                                if exp != 1 else input_features[ind]
                                for ind, exp in zip(inds, row[inds]))

removes the ambiguity.

Here's the above example run with the suggested change:

>>> from sklearn.preprocessing import PolynomialFeatures
>>> import pandas as pd 
>>> df = pd.DataFrame([[1, 2, 3], [4, 5, 6], [7, 8, 9]], columns=["A B", "C", "D"])
>>> df
   A B  C  D
0    1  2  3
1    4  5  6
2    7  8  9
>>> poly = PolynomialFeatures(2)
>>> poly.fit(df)
PolynomialFeatures(degree=2, include_bias=True, interaction_only=False)
>>> poly.transform(df)
array([[  1.,   1.,   2.,   3.,   1.,   2.,   3.,   4.,   6.,   9.],
       [  1.,   4.,   5.,   6.,  16.,  20.,  24.,  25.,  30.,  36.],
       [  1.,   7.,   8.,   9.,  49.,  56.,  63.,  64.,  72.,  81.]])
>>> poly.get_feature_names()
['1', 'x0', 'x1', 'x2', 'x0^2', 'x0*x1', 'x0*x2', 'x1^2', 'x1*x2', 'x2^2']
>>> poly.get_feature_names(input_features=df.columns.tolist())
['1', 'A B', 'C', 'D', 'A B^2', 'A B*C', 'A B*D', 'C^2', 'C*D', 'D^2']
>>> 

[jnothman]

that's not so much more readable!

[mohamed-ali]

@jnothman That might be true, but at least 'A B*C' adds a bit of readability compared to 'A B C'.
Can you think of something else that would be more readable?

I thought about these:

-'"A B"*"C"'
-'(A B)*(C)'

But I think they are ugly. We can also throw a warning and replace any whitespace within a feature name with underscore:

• 'A B C' => 'A_B C' or
• 'A B C' => 'A_B*C'

@amueller what do you think?

[jnothman]

In terms of backwards compatibility and avoiding cluttering the interface, I like the underscores. On the other hand, users need only modify the input in that case. We're really we're unlikely to get this perfect with strings in any case.

…

On 29 Mar 2018 5:04 pm, "Mohamed Ali Jamaoui" ***@***.***> wrote: @jnothman <https://github.com/jnothman> That might be true, but at least 'A B*C' adds a bit of readability compared to 'A B C'. Can you think of something else that would be more readable? I thought about these: -'"A B"*"C"' -'(A B)*(C)' But I think they are ugly. We can also throw a warning and replace any whitespace within a feature name with underscore: - 'A B C' => 'A_B C'or - 'A B C' => 'A_B*C' @amueller <https://github.com/amueller> what do you think? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#10742 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz65lPrjh51ADBmEQJwkuMUn_NJvZ6ks5tjHmKgaJpZM4SZMcL> .

[mohamed-ali]

@jnothman, I agree that there is no perfect solution to this, but I think we need at least to throw a warning in case the input_features contain a whitespace.

[mohamed-ali]

@amueller what do you think?

[euanmacinnes]

Add a custom separator='' parameter to the function, then the dev can make it what they want. I wouldn't bother with also processing the A^2 stuff, leave that as a formatting example afterwards. default is space for backwards compatibility