Skip to main content

NuSVR

NuSVR is a nu-parametrised support vector regressor. Rather than fixing the tube width epsilon, the nu parameter controls the fraction of support vectors and the algorithm learns epsilon from the data.

Algorithm​

The estimator runs a two-phase fit. The first pass solves the dual with epsilon = 0 to estimate the residual scale; the fitted tube width is then set to the 1 - nu quantile of the absolute residuals. The second pass re-fits with that effective epsilon. The dual is optimised by a projected sub-gradient method over the dual coefficients with a box bound derived from C. The prediction is the kernel-weighted sum of dual_coef_ plus the intercept.

Constructor​

Skigen::NuSVR<Scalar> model(
Scalar nu = 0.5,
Scalar C = 1.0,
Kernel kernel = Kernel::RBF,
int degree = 3,
Scalar gamma = 0,
Scalar coef0 = 0,
Scalar tol = 1e-3,
int max_iter = 1000,
std::optional<uint64_t> random_state = std::nullopt);

Parameters​

ParameterDefaultDescription
nu0.5Upper bound on the fraction of margin errors and lower bound on the fraction of support vectors, in (0, 1].
C1.0Regularisation strength; the box bound on the dual coefficients.
kernelRBFKernel function.
gamma0Kernel coefficient; 0 means 1 / (n_features * var).

Methods​

MethodDescription
fit(X, y)Fit the nu-SVM regressor; learns epsilon from nu.
predict(X)Predicted target values.
score(X, y)Coefficient of determination R².

Fitted Attributes​

AccessorDescription
support()Indices of the support vectors.
n_support()Number of support vectors.
epsilon_fitted()The tube width learned from nu.

Example​

#include <Skigen/SVM>

using K = Skigen::NuSVR<double>::Kernel;
Skigen::NuSVR<double> reg(/*nu=*/0.5, /*C=*/5.0, K::Linear);
reg.fit(X, y);
Eigen::VectorXd pred = reg.predict(X);

Parity Scope​

Mirrors the dense sklearn.svm.NuSVR API surface and the nu-to-epsilon mapping. The sub-gradient solver targets the same optimum as libsvm but is coarser, so support-vector counts and the fitted epsilon may differ. Sparse input is deferred.

API Reference

For full signatures see the NuSVR API Reference.