mgcv-rust 0.15.0

Rust implementation of Generalized Additive Models with Python bindings

mgcv_rust

A Rust port of R's mgcv package — Generalized Additive Models with automatic smoothing-parameter selection (REML / LAML) and PIRLS fitting — with first-class Python bindings.

The numerical core targets byte-for-byte parity with mgcv on common families, and the Python API is designed to feel like sklearn: a single Gam class, named-predictor DataFrame inputs, fit() / predict() / score(), posterior CI, and subset views for marginal analysis.

Latest release	0.11.0
Parity	554 / 0 / 0 on the mgcv comparison battery
Python tests	211 passed, 1 xfailed
Headline perf	2d_gaussian_additive_n50000_k15_cr — 394 ms → 97 ms (4.05× vs. mgcv)
Python wrapper	mgcv_rust.Gam (canonical), GAMFitter deprecated alias

Why another GAM library?

• R-equivalent answers, Python ergonomics. If you have R / mgcv users producing models and Python services consuming them, this gives both sides the same numbers from the same fit. The serialize() method produces a portable artefact for deployment.
• Fast. Aggregate time across 80 parity fixtures: 2.0 s. On the largest Gaussian fixture (n=50,000, k=15), mgcv_rust is ~4× faster than R's mgcv.
• Real CIs, paired diffs. predict_ci(X) returns (mean, lo, hi) so you never have to "+= intercept" by hand. predict_diff(from_X, to_X, level=...) gives a paired-posterior CI for the difference between two predictions — strictly tighter than the naive predict_ci difference.
• Frozen serving view. GamPredictor(gam) is an __slots__-locked, inference-only wrapper with strict input validation (feature_names_in_ enforcement) and a check_against(gam, X_sample) round-trip assertion for deployment safety.

Installation

The Python package is built from this repo via maturin:

# Clone, then in a venv:
pip install maturin
maturin develop --features python,blas,blas-system --release

After that, import mgcv_rust works.

For a Rust-only consumer, add to your Cargo.toml:

[dependencies]
mgcv_rust = { git = "https://github.com/AlekJaworski/nn_exploring" }

Quick start

import numpy as np
import pandas as pd
from mgcv_rust import Gam

rng = np.random.default_rng(0)
n = 500
X = pd.DataFrame({
    "x0": rng.uniform(-2, 2, n),
    "x1": rng.uniform(0, 5, n),
})
y = np.sin(X["x0"]) + 0.3 * (X["x1"] - 2.5)**2 + rng.normal(0, 0.1, n)

gam = Gam(family="gaussian").fit(X, y)

gam.predict(X[:5])                  # response-scale predictions
mean, lo, hi = gam.predict_ci(X[:5])  # 95% CI, response scale
gam.score(X, y)                     # adjusted R²
print(gam.summary())                # mgcv-style block

That's it for the simple case. Read on for the curated tour, or jump to docs/GETTING_STARTED.md for a step-by-step tutorial with worked examples.

---

Tour

1. Build the model

Gam() accepts pandas / polars / numpy. If you pass a DataFrame, the column names become the predictor names; otherwise you can pass them explicitly.

gam = Gam(
    family="gaussian",           # gaussian, binomial, poisson, gamma, tweedie, nb, t-dist, scat, ...
    link=None,                   # None → canonical link for the family
    k_default=10,                # default basis dimension per smooth (mgcv default)
    term_k_mapping={"x0": 25},   # per-predictor overrides
    method="REML",
)
gam.fit(X, y)

After fit, sklearn-style attributes are populated:

gam.coef_              # full β vector (intercept first)
gam.intercept_         # link-scale
gam.intercept_response_ # response-scale (= mean(y) for identity link)
gam.feature_names_in_  # predictor names, fitting order
gam.n_features_in_
gam.lambda_            # per-smooth smoothing parameter
gam.edf_               # per-smooth effective degrees of freedom
gam.k_, gam.bs_        # basis dimension / basis type per smooth
gam.vcov_              # posterior covariance of β̂

2. Predict on different scales

gam.predict(X)                       # response (default — same as mgcv predict(type="response"))
gam.predict(X, scale="link")         # linear predictor, no inverse link (mgcv type="link")

result = gam.predict(X, type="terms")  # decomposition
result.intercept                       # scalar, link scale
result.contributions                   # DataFrame (n × m_smooths), link-scale, centered on training
result.total                           # response-scale ≡ gam.predict(X)

Invariant: gam.predict(X) == inv_link(intercept + result.contributions.sum(axis=1)).

3. Confidence intervals

mean, lo, hi = gam.predict_ci(X, level=0.95)       # response, 95% CI
mean, lo, hi = gam.predict_ci(X, scale="link")     # link-scale CI
mean, lo, hi = gam.predict_ci(X, level=0.5)        # 50% CI (narrower)

mean matches gam.predict(X, scale=scale) exactly. The CI is built from n_samples (default 1000) draws of β ~ N(β̂, vcov).

The legacy predict_ci(X, alpha=0.05) → (lo, hi) form still works (emits a DeprecationWarning) so existing callers don't break.

4. Differences between predictions, with paired CI

# Per-row diff: to[i] - from[i]
diff = gam.predict_diff(from_X, to_X)

# With paired posterior CI — much narrower than predict_ci differences
diff, lo, hi = gam.predict_diff(from_X, to_X, level=0.95)

# One row broadcast against many
diff = gam.predict_diff(baseline_row, candidates, broadcast="from")
diff = gam.predict_diff(candidates, target_row, broadcast="to")

Identity-link only — for non-identity links the response-scale difference isn't linear in β, so the closed-form CI argument doesn't transfer.

5. Subset / marginal views

gam[name] returns a view that includes only the listed smooths (and optionally the intercept). All predict methods inherit it.

gam[["x0"]].predict(X)                              # marginal effect of x0, response scale
gam[["x0"]].predict(X, scale="deviation")           # link-scale, intercept zeroed (sum-to-zero on train)
gam[["x0", Gam.INTERCEPT]].predict(X)               # marginal effect + intercept
gam[["x0"]].partial_effect("x0").plot()             # what the smooth looks like

All sklearn attributes filter to the active features:

gam[["x0"]].coef_            # just the x0 block (+ intercept if active)
gam[["x0"]].feature_names_in_  # ["x0"]

6. Plotting, summary, score

gam.plot()                  # figure with one subplot per smooth (CI bands shaded)
gam.plot("x0")              # single-smooth axes
df = gam.partial_effect("x0", level=0.95)  # underlying data: columns x0, effect, lo, hi

print(gam.summary())
# Gam summary  family=gaussian  link=identity  n_obs=500
#   intercept (link)     = -0.0123
#   intercept (response) = -0.0123
#   scale (σ²)           = 0.0101
#   deviance             = 4.8732
#   R² (adj)             = 0.9612
#   smooths:
#     s(          x0)  k= 10  edf=  6.13  λ=2.345e-02
#     s(          x1)  k= 10  edf=  3.07  λ=4.890e-01

gam.score(X, y)              # adjusted R² for regression; accuracy@0.5 for binomial
gam.predict_proba(X)         # binomial only: (n, 2) [[P(0), P(1)]]

7. Serving with GamPredictor

For deployment paths, wrap the fitted Gam in a GamPredictor. Same API surface, plus strict column validation and a round-trip assertion.

from mgcv_rust import GamPredictor

predictor = GamPredictor(gam)
predictor.check_against(gam, X[:50])   # raises if predictions diverge
predictor.predict(X_serve)             # ValueError if any expected column is missing
predictor[["x0"]].predict(X_serve)     # subset view returns another GamPredictor

# `__slots__` blocks attribute leaks — once built, the bound Gam's
# coef/vcov/schema are the contract.

This structurally closes two production bug classes:

• Column index drift when the serving DataFrame's columns differ from training: feature_names_in_ is enforced on every predict call (re-ordered → realigned; missing → ValueError).
• x-grid clamping in interpolation-based predictors: GamPredictor recomputes the basis at the requested X via evaluate_lpmatrix, so there's no pre-computed grid to clamp against.

Families and links

Family	Default link	Other links	Notes
gaussian	identity	—	bs-spline / cubic-regression / random-effects bases supported
binomial	logit	—	predict_proba() available
poisson	log	—
gamma	inverse	log	log-link via link="log"
tweedie	log	—	tweedie_p= (default 1.5); fixed-p
nb / negbin	log	—	nb profiles θ; negbin is fixed-θ
t-dist / scat	identity	—	df profiled if not given (df ∈ [2, 100])
quasipoisson, quasibinomial	log / logit	—	dispersion-aware
quantile	identity	—	see mgcv_rust.fit_quantile / fit_quantile_lss

Architecture (Rust side)

File	Responsibility
src/gam.rs	GAM struct + fit / predict entry points
src/pirls.rs	Penalized IRLS inner loop
src/reml/	Outer-loop REML / LAML optimization
src/smooth.rs	Basis functions (cubic-regression, B-spline, random effects, tensor products)
src/penalty.rs	Penalty-matrix construction
src/lib.rs	PyO3 bindings — PyGAM, compute_penalty_matrix, newton_pirls_py

Build features:

• python — enable PyO3 bindings.
• blas / blas-system — link against system BLAS for matmul-heavy paths.

Parity and performance

Run the parity battery against R/mgcv:

pytest tests/parity/ -q
# 554 passed, 0 failed, 0 xfailed, 0 skipped

Microbench:

python3 scripts/python/bench_step_blend.py 5

Headline (2d_gaussian_additive_n50000_k15_cr, identity link):

	Time
R mgcv	~394 ms
mgcv_rust 0.11.0	97 ms

Status and limitations

• Joint (ρ, log φ) outer Newton not yet implemented — the binding constraint for closing the remaining performance gap on dispersion-bearing GLMs. Tracked in mgcv_rust - Backlog - Next Numerical Steps (Obsidian).
• predict_diff is identity-link only; non-identity raises with a workaround pointing at get_posterior_samples.
• Auto-k tuning is opt-in via Gam(auto_k=True). Default is a single fit at k_default=10 (mgcv's default), with term_k_mapping overrides — closer to mgcv's "tune k, run k.check" convention and avoids hidden multi-fit costs.
• sklearn BaseEstimator mixin (for Pipeline / GridSearchCV) is not yet wrapped — soft-dep, deferred.

References

• Wood, S.N. (2017). Generalized Additive Models: An Introduction with R (2nd ed.). Chapman and Hall/CRC.
• Wood, S.N. (2011). Fast stable restricted maximum likelihood and marginal likelihood estimation of semiparametric generalized linear models. J. R. Stat. Soc. B, 73(1), 3–36.

License

MIT — see LICENSE.