Module: aprender::glm

Public module of the aprender-core crate.

Source

crates/aprender-core/src/glm.rs or directory.

Example

use aprender::glm::{GLM, Family, Link};
// See `cargo doc -p aprender-core --open` for full API reference.

Module summary

aprender::glm provides generalized linear models — the natural extension of linear regression to non-Gaussian response distributions. A single GLM struct is parameterised by a Family enum (Poisson, Negative Binomial, Gamma, Binomial, Gaussian, …) and an optional Link function (Log, Inverse, Logit, Identity). Fitting uses iteratively reweighted least squares (IRLS). Reach for GLM when you have count data (Poisson / NegBinomial), positive continuous outcomes (Gamma), proportions (Binomial), or any other exponential family.

Key types

Usage patterns

Pattern 1: Poisson regression for count data

use aprender::glm::{GLM, Family, Link};
use aprender::primitives::{Matrix, Vector};

let x = Matrix::from_vec(5, 1, vec![1.0, 2.0, 3.0, 4.0, 5.0]).expect("5x1");
let y = Vector::from_slice(&[1.0, 3.0, 5.0, 8.0, 13.0]);

let mut glm = GLM::new(Family::Poisson)
    .with_link(Link::Log)
    .with_max_iter(100)
    .with_tolerance(1e-6);

glm.fit(&x, &y).expect("Poisson GLM fit");

let test = Matrix::from_vec(1, 1, vec![6.0]).expect("1x1");
let pred = glm.predict(&test).expect("predict");
println!("expected count at x=6 ≈ {:.2}", pred[0]);

Pattern 2: Gamma regression for positive continuous data

use aprender::glm::{GLM, Family, Link};
use aprender::primitives::{Matrix, Vector};

let x = Matrix::from_vec(5, 1, vec![1.0, 2.0, 3.0, 4.0, 5.0]).expect("5x1");
// strictly positive y (e.g. insurance claim amounts)
let y = Vector::from_slice(&[1.2, 2.5, 5.1, 9.8, 19.5]);

let mut glm = GLM::new(Family::Gamma)
    .with_link(Link::Log)
    .with_dispersion(0.1);

glm.fit(&x, &y).expect("Gamma GLM fit");

let coefs = glm.coefficients().expect("fitted");
println!("coefficients: {:?}", coefs);

See also

• linear_model — Gaussian-noise linear regression (a special case of GLM)
• bayesian — Bayesian counterparts via conjugate priors
• classification — logistic regression is a special Binomial-family GLM
• metrics — deviance and other GLM-specific scores

Full API

Run cargo doc -p aprender-core --open for the rendered rustdoc, or browse docs.rs/aprender for the published version.