Generalized Additive Models in Python (pyGAM)

Warnings

• gotcha Installing NumPy/SciPy linked to Intel MKL for acceleration can be tricky, especially with Conda due to channel compatibility issues. Pip's NumPy-MKL is often outdated. Consider third-party builds or specific Conda channels if MKL optimization is critical.
  Fix: For MKL acceleration, explicitly search for and install NumPy/SciPy with MKL via specific Conda channels (e.g., intel channel) or use third-party builds (e.g., `pip install numpy scipy --extra-index-url https://urob.github.io/numpy-mkl`).
• gotcha P-values derived from models where smoothing parameters have been estimated may be lower than they should be, leading to an increased rate of false positives (rejecting the null hypothesis too readily).
  Fix: Interpret p-values with caution, especially when smoothing parameters are estimated. Consider alternative model evaluation metrics or sensitivity analyses.
• gotcha When combining a spline term (e.g., `s(feature)`) and a linear term for the *same* feature in a GAM, it can introduce a model identifiability problem. This can cause p-values to appear statistically significant when they are not.
  Fix: Avoid including both a spline and a linear term for the same feature in a single GAM specification. Use either `s()` for a flexible non-linear effect or `l()` for a purely linear effect.
• gotcha The `pyGAM` package available on `conda-forge` is typically less up-to-date than the version available via `pip`.
  Fix: For the most current features and bug fixes, prioritize installing `pygam` using `pip`.
• gotcha For large models with constraints, installing `scikit-sparse` can significantly improve optimization performance due to its faster sparse Cholesky factorization. It also has a dependency on `nose` for its import.
  Fix: If working with large or constrained models, install `scikit-sparse` and `nose`: `conda install -c conda-forge scikit-sparse nose` (or equivalent pip install if available).
• gotcha pyGAM is officially tested with Python 3.10+ and is compatible up to Python 3.13. Using it with significantly older or very new, untested Python versions might lead to unexpected compatibility issues.
  Fix: Ensure your Python environment is within the supported range (3.10 to 3.13) for optimal compatibility.

Install

• pip install pygam Install stable version

Imports

• LinearGAM

  from pygam import LinearGAM

  Primary class for fitting Linear Generalized Additive Models
• LogisticGAM

  from pygam import LogisticGAM

  Primary class for fitting Logistic Generalized Additive Models
• PoissonGAM

  from pygam import PoissonGAM

  Primary class for fitting Poisson Generalized Additive Models
• s

  from pygam import s

  Helper function to define a spline term for a feature
• f

  from pygam import f

  Helper function to define a factor term for a categorical feature
• te

  from pygam import te

  Helper function to define a tensor product interaction term

Quickstart

This quickstart demonstrates how to install pyGAM, load an example dataset, define a LinearGAM with spline and factor terms, fit the model to the data, and then print a statistical summary. It also includes a basic prediction example.

import numpy as np
from pygam import LinearGAM, s, f
from pygam.datasets import wage

X, y = wage() # Load example data

# Define a GAM with a spline term for features 0 and 1, and a factor term for feature 2
gam = LinearGAM(s(0) + s(1) + f(2))

# Fit the model
gam.fit(X, y)

# Print a summary of the model fit
print(gam.summary())

# Example of predicting (using dummy data for simplicity)
dummy_X = np.array([[10, 20, 1], [15, 25, 0]])
predictions = gam.predict(dummy_X)
print(f"Predictions for dummy data: {predictions}")