High-performance machine learning for modern C++ and Eigen.

   
 

---

About

Skigen is a header-only C++ template library for machine learning, built on Eigen. It brings the scikit-learn API — fit(), transform(), predict() — to native C++.

• Skigen is versatile. Preprocessing, linear & Bayesian linear models, decomposition, clustering, trees, ensembles (random forests, gradient boosting, histogram GB), naive Bayes, SVMs (linear + kernel SMO), MLPs, manifold learning (Isomap/MDS/LLE/SpectralEmbedding/t-SNE/UMAP), feature selection, calibration, isotonic regression, hyperparameter search (Grid/Randomized), pipelines, metrics — covering the scikit-learn 1.7.x surface with a consistent API.
• Skigen is fast. Eigen's expression templates, explicit SIMD vectorization, and compile-time polymorphism via CRTP. No interpreter, no garbage collector, no runtime dispatch.
• Skigen is elegant. Header-only — drop Skigen/ next to Eigen/ and #include. The same fit / transform / predict workflow, native to modern C++.

Design

Principle	Implementation
Eigen-native	Headers, namespaces, and include patterns mirror Eigen. #include <Skigen/Dense> feels like #include <Eigen/Dense>.
Header-only	Drop Skigen/ next to Eigen/ — no compiled libraries, no linker flags.
Templatized	All estimators accept Scalar (default: double). Switch to float for 2× SIMD throughput on the same hardware.
Zero-copy	Inputs use Eigen::Ref<const MatrixType> — supports sub-blocks and memory-mapped data without copying.
CRTP	Static polymorphism via the Curiously Recurring Template Pattern. Zero vtable overhead.
Bit-level parity	Results match scikit-learn for identical inputs and default parameters. Verified via cross-language parity tests.

Quick Start

#include <Eigen/Dense>
#include <Skigen/Dense>
#include <iostream>

int main() {
    Eigen::MatrixXd X(4, 2);
    X << 1, 10,
         2, 20,
         3, 30,
         4, 40;

    Skigen::StandardScaler scaler;
    Eigen::MatrixXd Z = scaler.fit_transform(X);

    std::cout << "Standardized:\n" << Z << "\n";

    // Round-trip back to original scale
    Eigen::MatrixXd X_back = scaler.inverse_transform(Z);
    std::cout << "Recovered:\n" << X_back << "\n";

    return 0;
}

Build & Test

git clone https://github.com/skigen-project/skigen.git
cd skigen

cmake -B build -DSKIGEN_BUILD_TESTS=ON
cmake --build build
./build/tests/skigen_tests

Eigen is detected automatically if installed system-wide, or from a sibling ../eigen source tree.

Include Pattern

Skigen follows the same convention as Eigen — module headers without file extensions:

#include <Skigen/Core>           // Base classes, traits, concepts
#include <Skigen/Preprocessing>  // Scalers, normalizers, ...
#include <Skigen/Dense>          // Everything bundled

Repository Layout

Skigen/                     # Header library (Eigen-style)
├── Core                    # Module header — base classes, traits, concepts
├── Preprocessing           # Module header — scalers, normalizers, ...
├── LinearModel             # Module header — regression, classification, Bayesian
├── Decomposition           # Module header — PCA, TruncatedSVD
├── Cluster                 # Module header — KMeans, MiniBatchKMeans
├── Neighbors               # Module header — KNN classifier/regressor
├── Tree                    # Module header — decision trees
├── Ensemble                # Module header — RandomForest, (Hist)GradientBoosting
├── NaiveBayes              # Module header — Gaussian/Multinomial/Bernoulli NB
├── SVM                     # Module header — Linear SVM, kernel SVM (SMO)
├── NeuralNetwork           # Module header — MLPClassifier, MLPRegressor
├── FeatureSelection        # Module header — VarianceThreshold, SelectKBest, ...
├── Manifold                # Module header — Isomap, MDS, LLE, SE, t-SNE, UMAP
├── Calibration             # Module header — CalibratedClassifierCV
├── Isotonic                # Module header — IsotonicRegression
├── ModelSelection          # Module header — split/CV, Grid/RandomizedSearchCV
├── Pipeline                # Module header — compile-time pipeline composition
├── Metrics                 # Module header — regression, classification, pairwise
├── Dense                   # Convenience header — bundles all modules
└── src/                    # Internal headers (.h)
    ├── Core/               # Traits, Concepts, Base, Validation, EigenHelpers
    ├── Preprocessing/      # StandardScaler, MinMaxScaler, MaxAbsScaler, ...
    ├── LinearModel/        # LinearRegression, Ridge, Lasso, BayesianRidge, ARD
    ├── Decomposition/      # PCA, TruncatedSVD
    ├── Cluster/            # KMeans, MiniBatchKMeans
    ├── Neighbors/          # KNeighborsClassifier, KNeighborsRegressor
    ├── Tree/               # DecisionTreeClassifier, DecisionTreeRegressor
    ├── Ensemble/           # RandomForest, GradientBoosting, HistGradientBoosting
    ├── NaiveBayes/         # GaussianNB, MultinomialNB, BernoulliNB
    ├── SVM/                # LinearSVC, LinearSVR, SVC, SVR, Nu*/OneClassSVM
    ├── NeuralNetwork/      # MLPClassifier, MLPRegressor
    ├── FeatureSelection/   # VarianceThreshold, SelectKBest, SelectFromModel, RFE
    ├── Manifold/           # Isomap, MDS, LLE, SpectralEmbedding, TSNE, UMAP
    ├── Calibration/        # CalibratedClassifierCV
    ├── Isotonic/           # IsotonicRegression
    ├── ModelSelection/     # TrainTestSplit, CrossValidation, Grid/RandomizedSearchCV
    ├── Pipeline/           # Pipeline (compile-time)
    └── Metrics/            # Regression, Classification, Pairwise
tests/                      # Unit + parity tests
benchmarks/                 # Performance benchmarks
examples/                   # Usage examples (one per public estimator)
doc/                        # Requirements + Docusaurus website

Requirements

	Minimum
C++ Standard	Latest (currently C++23)
Compilers	GCC ≥ 13, Clang ≥ 17, MSVC ≥ 19.38
Eigen	≥ 3.4
CMake	≥ 3.20 (for building tests/examples)

License

Skigen is released under the MIT License.

Acknowledgments

Skigen builds on Eigen for numerical computation and adopts the API conventions established by scikit-learn.