Constrained Logistic Regression

Sample implementation of constructing a logistic regression with given ranges on each of the feature's coefficients (via clogistic library).

The Data

We will use the processed version of telco customer churn data from Kaggle. The data can be downloaded here.

Steps

Define the constraints

For example:

# define constraints as dataframe
import numpy as np
constraint_df = pd.DataFrame(data=[
                                   ['gender',-np.inf,np.inf],
                                   ['SeniorCitizen',-np.inf,np.inf],
                                   ['Partner',-np.inf, 0],
                                   ['Dependents',-np.inf,0],
                                   ['tenure',-np.inf,0],
                                   ['PhoneService',-np.inf,0],
                                   ['PaperlessBilling',-np.inf,np.inf],
                                   ['MonthlyCharges',-np.inf,np.inf],
                                   ['intercept',-np.inf,np.inf]],
                             columns=['feature','lower_bound','upper_bound'])
constraint_df

|    | feature          |   lower_bound |   upper_bound |
|---:|:-----------------|--------------:|--------------:|
|  0 | gender           |          -inf |           inf |
|  1 | SeniorCitizen    |          -inf |           inf |
|  2 | Partner          |          -inf |             0 |
|  3 | Dependents       |          -inf |             0 |
|  4 | tenure           |          -inf |             0 |
|  5 | PhoneService     |          -inf |             0 |
|  6 | PaperlessBilling |          -inf |           inf |
|  7 | MonthlyCharges   |          -inf |           inf |
|  8 | intercept        |          -inf |           inf |

Model training via clogistic

# train using clogistic
from scipy.optimize import Bounds
from clogistic import LogisticRegression as clLogisticRegression

lower_bounds = constraint_df['lower_bound'].to_numpy()
upper_bounds = constraint_df['upper_bound'].to_numpy()
bounds = Bounds(lower_bounds, upper_bounds)

cl_logreg = clLogisticRegression(penalty='none')
cl_logreg.fit(X_train, y_train, bounds=bounds)

Retrieve the model coefficients

# coefficients as dataframe
cl_coef = pd.DataFrame({
    'feature': df.drop(columns='Churn').columns.tolist() + ['intercept'],
    'coefficient': list(cl_logreg.coef_[0]) + [cl_logreg.intercept_[0]]
})

cl_coef

|    | feature          |   coefficient |
|---:|:-----------------|--------------:|
|  0 | gender           |   0.0184168   |
|  1 | SeniorCitizen    |   0.506692    |
|  2 | Partner          |   3.85603e-09 |
|  3 | Dependents       |  -0.35721     |
|  4 | tenure           |  -0.0557211   |
|  5 | PhoneService     |  -0.796233    |
|  6 | PaperlessBilling |   0.398824    |
|  7 | MonthlyCharges   |   0.033197    |
|  8 | intercept        |  -1.36086     |