gamdist#

gamdist fits the GLM/GAM zoo — binary, continuous, or count outcomes paired with continuous, categorical, or spline-transformed features, with arbitrary convex regularization attached to whichever terms want it. Every such model is a single convex optimization problem.

The library provides:

  • Feature types: linear (ridge), categorical (L1, L2, group lasso, network lasso, network ridge), and spline with an integrated curvature penalty

  • Families and links: normal, binomial, Poisson, gamma, exponential, and inverse Gaussian; identity, logit, probit, complementary log-log, log, reciprocal, and reciprocal-squared links

  • Multi-task fitting: MultiTaskGAM for K correlated responses with optional cross-task coupling

  • Model selection: AIC, AICc, GCV, UBRE, and effective degrees of freedom for each feature

The ADMM decomposition of [CKB13] makes the zoo tractable by splitting the joint problem into a per-feature primal step plus a per-outcome proximal step coordinated by dual variables. Parallelism is a side effect; the real prize is modularity — outcomes, features, and regularizers are independent components that mix and match in any combination without any side needing to know about the others. Full citations are on the References page.

Quickstart#

The core workflow: construct a GAM, register features with add_feature(), then call fit(). For binary outcomes, change family to "binomial" and the rest of the API stays the same.

import numpy as np
import pandas as pd
from gamdist import GAM

X = pd.DataFrame(
    {
        "purchases": np.random.choice([0, 3, 10, 16], size=1000),
        "gender": np.random.choice(["male", "female"], size=1000),
    }
)
y = (
    0.1 * np.log1p(X["purchases"].values)
    + np.where(X["gender"].values == "male", 0.1, -0.5)
    + np.random.normal(size=1000) * 0.1
)

mdl = GAM(family="normal")
mdl.add_feature(name="purchases", type="linear", transform=np.log1p)
mdl.add_feature(name="gender", type="categorical")
mdl.fit(X, y)

mdl.summary()
yhat = mdl.predict(X)

The methods in this library follow the algorithm of [CKB13], built on the ADMM framework of [BPC11].

Contents:

Indices and tables#