StudentTLikelihood<T>

Models & Types · AiDotNet.GaussianProcesses

Implements the Student-t likelihood for robust regression.

Implements: ILikelihood<T>

For Beginners

The Student-t likelihood provides robust regression that is less sensitive to outliers than Gaussian likelihood.

y = f + ε, where ε ~ Student-t(0, σ², ν)

The degrees of freedom ν controls “heaviness” of tails:

  • ν → ∞: Approaches Gaussian (not robust)
  • ν = 4-5: Moderately robust
  • ν = 1: Very robust (Cauchy distribution)

When to use:

  • Data has outliers
  • Errors might not be Gaussian
  • You want predictions to not be pulled by extreme values

Note: Student-t likelihood requires approximate inference.

Constructors

Constructor Summary
StudentTLikelihood(double noiseScale, double degreesOfFreedom) Initializes a new Student-t likelihood.

Properties

Property Summary
double DegreesOfFreedom { get; } Gets the degrees of freedom.
double NoiseScale { get; } Gets the noise scale.
string Name { get; } Gets the name of this likelihood.

Methods

Method Summary
T LogLikelihood(Vector<T> y, Vector<T> f) Computes the log-likelihood of observations.
T PredictiveVariance(T fMean, T fVariance) Computes predictive variance.
T TransformMean(T f) Transforms latent function value (identity for Student-t).
Vector<T> LogLikelihoodGradient(Vector<T> y, Vector<T> f) Computes the gradient of the log-likelihood.
Vector<T> LogLikelihoodHessianDiag(Vector<T> y, Vector<T> f) Computes the Hessian diagonal of the log-likelihood.