Skip to content

implemented-methods.md

Latest commit

 

History

History
161 lines (126 loc) · 5.84 KB

implemented-methods.md

File metadata and controls

161 lines (126 loc) · 5.84 KB
Raw

Implemented Methods

Last updated: 2026-07-24
Switch: Chinese

This page is the maintained inventory of public models, functions, and major solver families in statgpu. Detailed mathematical and backend contracts live on the linked model and guide pages.

Regression and Generalized Linear Models

Class Description Backends
LinearRegression OLS with classical, HC0–HC3, and HAC inference NumPy, CuPy, Torch
Ridge L2-penalized linear regression NumPy, CuPy, Torch
Lasso L1 regression with debiased/bootstrap inference paths NumPy, CuPy, Torch
ElasticNet L1+L2 penalized regression NumPy, CuPy, Torch
LogisticRegression Binary logistic/probit regression NumPy, CuPy, Torch
PoissonRegression Poisson GLM NumPy, CuPy, Torch
GammaRegression Gamma GLM NumPy, CuPy, Torch
InverseGaussianRegression Inverse Gaussian GLM NumPy, CuPy, Torch
NegativeBinomialRegression Negative-binomial GLM NumPy, CuPy, Torch
TweedieRegression Tweedie GLM NumPy, CuPy, Torch
QuantileRegression Quantile regression with kernel/bootstrap inference NumPy, CuPy, Torch
OrderedLogitRegression Ordered logit with analytical-Hessian inference NumPy, CuPy, Torch
OrderedProbitRegression Ordered probit with analytical-Hessian inference NumPy, CuPy, Torch

Penalized Models

The penalty registry includes L1, L2, Elastic Net, SCAD, MCP, adaptive L1, group Lasso, adaptive group Lasso, group MCP, and group SCAD implementations. Aliases are accepted for selected penalties; the registry and compatibility matrix are the source of truth rather than a hard-coded count.

Class Loss or model family Backends
PenalizedGeneralizedLinearModel Unified penalized GLM interface NumPy, CuPy, Torch
PenalizedLinearRegression Penalized Gaussian regression NumPy, CuPy, Torch
PenalizedLogisticRegression Penalized binary regression NumPy, CuPy, Torch
PenalizedPoissonRegression Penalized Poisson regression NumPy, CuPy, Torch
PenalizedQuantileRegression Quantile loss with proximal/FISTA paths NumPy, CuPy, Torch
PenalizedRobustRegression Huber, bisquare, and fair losses where supported NumPy, CuPy, Torch
PenalizedCoxPHModel Penalized Cox partial likelihood NumPy, CuPy, Torch

Solver availability depends on the selected loss and penalty. Consult the Loss × Penalty × Solver Framework and Solver × Penalty Matrix before choosing an explicit solver.

Example

from statgpu.linear_model import PenalizedGeneralizedLinearModel

# L1 is non-smooth, so use FISTA or solver="auto".
model = PenalizedGeneralizedLinearModel(
    loss="poisson",
    penalty="l1",
    alpha=0.05,
    solver="fista",
)
model.fit(X, y)

Cross-Validation

Class Description Backends
RidgeCV Ridge alpha selection NumPy, CuPy, Torch
LassoCV Warm-start Lasso path NumPy, CuPy, Torch
ElasticNetCV Joint l1_ratio and alpha search NumPy, CuPy, Torch
LogisticRegressionCV Logistic-regression CV NumPy, CuPy, Torch
PenalizedGLM_CV Unified penalized-GLM CV NumPy, CuPy, Torch
CoxPHCV Cox penalty search and final refit NumPy, CuPy; see CoxPH docs

ANOVA

  • f_oneway
  • f_twoway
  • f_welch
  • tukey_hsd
  • bonferroni
  • cohens_f
  • partial_eta_squared

See ANOVA for design restrictions and scalar distribution boundaries.

Covariance Estimation

  • EmpiricalCovariance
  • LedoitWolf
  • OAS
  • ShrunkCovariance
  • MinCovDet
  • GraphicalLasso
  • GraphicalLassoCV

See Covariance Estimation.

Panel Data

  • PanelOLS
  • RandomEffects
  • PooledOLS
  • BetweenOLS
  • FirstDifferenceOLS
  • FamaMacBeth

See Panel Data Models for covariance, rank-deficiency, and backend-preserving prediction contracts.

Nonparametric and Semiparametric Methods

  • KernelDensity and kernel regression
  • KernelRidge and KernelRidgeCV
  • KernelPCA
  • Nystroem
  • SplineTransformer
  • B-spline, natural cubic, cyclic cubic, and thin-plate spline bases
  • GAM

Unsupervised Learning

  • PCA, TruncatedSVD, IncrementalPCA
  • NMF, MiniBatchNMF
  • KMeans, MiniBatchKMeans, DBSCAN
  • GaussianMixture, AgglomerativeClustering
  • UMAP, TSNE, NNDescent

Survival Analysis

Class Description Backends
CoxPH Breslow/Efron/Exact ties, delayed entry, (start, stop] rows, strata, robust/cluster inference, backend-native prediction NumPy, CuPy, Torch
CoxPHCV L2 grid selection with the same risk-set semantics and subject-preserving folds NumPy, CuPy, Torch
PenalizedCoxPHModel Standard right-censored Cox partial likelihood with convex/non-convex penalties where supported NumPy, CuPy, Torch

The base installation contains the maintained Cox implementation. The optional statgpu[survival] extra installs statsmodels for external validation and comparison; it is not required for delayed entry or strict Breslow/Efron robust inference. See Cox Proportional Hazards for the precise support matrix.

Feature Selection and Diagnostics

  • StepwiseSelector and stepwise_selection
  • fixed-X and model-X knockoff filters and selector wrappers
  • RegressionDiagnostics and diagnose_model

Multiple Testing and Resampling

  • adjust_pvalues
  • combine_pvalues
  • permutation_test
  • bootstrap utilities exposed by the inference API

Validation Scope

Backend support in this inventory means the public execution path exists. Numerical, performance, and physical-GPU claims remain scoped to the exact model, backend, hardware, and commit recorded by the corresponding tests or validation artifact.