How to interpret the `coef_` attribute from a fitted model?

[jolespin]

I am trying to understand the tutorial: https://codeburst.io/pygam-getting-started-with-generalized-additive-models-in-python-457df5b4705f

How can I interpret the coef_? Why are there 121 coefficients with 6 attributes?

import pandas as pd        
from pygam import LogisticGAM
from sklearn.datasets import load_breast_cancer

#load the breast cancer data set
data = load_breast_cancer()
#keep first 6 features only
df = pd.DataFrame(data.data, columns=data.feature_names)[['mean radius', 'mean texture', 'mean perimeter', 'mean area','mean smoothness', 'mean compactness']]
y = pd.Series(data.target)
X = df[['mean radius', 'mean texture', 'mean perimeter', 'mean area','mean smoothness', 'mean compactness']]

#Fit a model with the default parameters
gam = LogisticGAM().fit(X, y)
X.shape, gam.coef_.shape
# ((569, 6), (121,))

[daventero]

By default pygam creates 20 splines per feature, that is 120 coefficients...one per each spline in each feature, plus the intercept....121.

You can check the source code here https://github.com/dswah/pyGAM/blob/master/pygam/terms.py#L1639

Does that help you?

[daventero]

Actually you could change the number of splines in each feature via the n_splines argument.

from pygam import LinearGAM, s, f
from pygam.datasets import wage

X, y = wage()
gam = LinearGAM(s(0, n_splines=30) + s(1, n_splines=30) + f(2)).fit(X, y)

print(X.shape, gam.coef_.shape)
# ((3000, 3), (66,))

As you can see there are , 30 coefficients, per each 'splined' feature plus, 1 for the intercept, plus 5 coefficients for the categorical (or factor) variable, one for each level, as f() function performs one-hot-encoding.

[jolespin]

I will get back to you on this after I read this. I have never heard about GAMs until I watched the PyData talk and I am extremely interested. I need to figure out how to interpret all of the spline values for an independent variable (covariate?) / attribute. Thanks for your response. I look forward in diving deep into these methods.

[dswah]

@jolespin i will close this issue. please re-open it as you see fit.