如何在python中的数据框中快速生成二次数值特征？

Question

Using python and standard libraries I'd like to quickly generate interaction features for machine learning models (classifiers or regressors). Because feature engineering by hand can be time consuming I'm looking for standard python libraries and methods that can semi-automate some of the process. For example, to generate quadratic features for analysis I have the following code:

import pandas as pd
import numpy as np

df = pd.DataFrame({'a': ['abc', 'def', 'ghi', 'kjl'],
                   'b': [2, 5, 7, 8],
                   'c': [1.2, 3, 4, 6]})
num_cols = [col for col in df.columns if df[col].dtype in [np.int64, np.float64]]
quadratic_cols = [tuple(sorted((i,j))) for i in num_cols for j in num_cols]
quad_col_pairs = list(set(quadratic_cols))

for col_pair in quad_col_pairs:
    col1, col2 = col_pair
    quadratic_col = '{}*{}'.format(*col_pair)
    df[quadratic_col] = df[col1] * df[col2]

I'd like to simplify this code because this kind of feature engineering should more standardized and quickly deployed. It also falls short because it would require more lines of code to generate derived features from addition, subtraction, or division across the feature columns.

How can I simplify the above code? Is there a standard python method or library that can more efficiently generate derived features for building models?

Answer 1

Try this for getting required columns with avoiding loops,

import itertools
L=df.select_dtypes(include=[np.number]).columns.tolist()
quad_col_pairs =  list(itertools.combinations_with_replacement(L,2))

for col_pair in quad_col_pairs:
    col1, col2 = col_pair
    quadratic_col = '{}*{}'.format(*col_pair)
    df[quadratic_col] = df[col1] * df[col2]

Answer 2

Since you explicitly tag it with scikit-learn: you can use PolynomialFeatures :

from sklearn.preprocessing import PolynomialFeatures
pf = PolynomialFeatures(include_bias=False)
pf.fit_transform(df._get_numeric_data()) 

#array([[ 2.  ,  1.2 ,  4.  ,  2.4 ,  1.44],
#       [ 5.  ,  3.  , 25.  , 15.  ,  9.  ],
#       [ 7.  ,  4.  , 49.  , 28.  , 16.  ],
#       [ 8.  ,  6.  , 64.  , 48.  , 36.  ]])

It also gives you options to use higher order polynomials, and to include only the interaction terms.