ProductStructureKernel<T>

Models & Types · AiDotNet.Kernels

Product Structure Kernel for modeling multiplicative interactions between feature groups.

Implements: IKernelFunction<T>

For Beginners

The Product Structure Kernel assumes the underlying function has a multiplicative structure: f(x) = f_1(x_G1) × f_2(x_G2) × … × f_K(x_GK)

Where x_Gi is the subset of features in group i.

The kernel is: k(x, x’) = k_1(x_G1, x’_G1) × k_2(x_G2, x’_G2) × … × k_K(x_GK, x’_GK)

This is useful when:

  • Features naturally group together (e.g., spatial × temporal)
  • There’s known multiplicative interaction structure
  • You want to reduce computation via Kronecker structure

Example: Modeling sales as (location effects) × (time effects) × (product effects)

Compare to Additive Structure:

  • Additive: f = f_1 + f_2 + … (sum of independent effects)
  • Product: f = f_1 × f_2 × … (multiplicative interaction)

Product kernels can capture stronger interactions but are less interpretable.

Constructors

Constructor Summary
ProductStructureKernel(IKernelFunction<T>[] groupKernels, int[][] featureGroups) Initializes a Product Structure Kernel.

Properties

Property Summary
int NumGroups { get; } Gets the number of feature groups.
int[][] FeatureGroups { get; } Gets the feature indices for each group.

Methods

Method Summary
Matrix<T>[] ComputeGroupKernelMatrices(Matrix<T>[] Xgroups) Computes the kernel matrix with Kronecker structure (for gridded data).
ProductStructureKernel<T> FullyFactorized(int numFeatures, double baseLengthscale) Creates a Product Structure Kernel with automatic grouping (one feature per group).
ProductStructureKernel<T> WithRBF(int[][] featureGroups, double[] lengthscales) Creates a Product Structure Kernel with RBF for all groups.
T Calculate(Vector<T> x1, Vector<T> x2) Calculates the product kernel value.
T LogDeterminant(Matrix<T>[] groupKernelMatrices) Computes log-determinant using Kronecker structure.
Vector<T> KroneckerMultiply(Matrix<T>[] groupKernelMatrices, Vector<T> v) Performs efficient Kronecker matrix-vector product.