AI Glossary
The complete dictionary of Artificial Intelligence
Multiple Kernel SVM
A variant of Support Vector Machines that combines multiple kernel functions, often weighted, to improve data separation by simultaneously capturing different feature structures or scales.
Linear Combination of Kernels
A kernel aggregation method where the final kernel is a weighted sum of base kernels, allowing the fusion of their respective representations in the implicit feature space.
Multiplicative Kernel
A kernel function constructed by the product of several kernels, favoring feature intersection and strengthening similarity only when individual components are simultaneously similar.
Kernel Weight Learning
An optimization process that automatically determines the optimal weighting coefficients for each kernel in a combination, typically integrated into the SVM's cost function.
MKL (Multiple Kernel Learning)
An algorithmic framework that simultaneously learns the SVM classifier and the optimal kernel combination, treating kernel weights as additional parameters to be optimized.
Heterogeneous Kernel
A kernel designed to operate on different data types (numerical, categorical, textual) by combining kernels specific to each data view or modality.
Block Coordinate Descent
An alternating optimization algorithm used in MKL, which sequentially updates kernel weights and classifier parameters by fixing some to optimize others, ensuring convergence.
Kernel Regularization
A model complexity control technique in multiple kernel SVMs that penalizes excessive kernel weights to prevent overfitting and promote sparse combinations.
Variable Bandwidth Kernel
Kernel function (often RBF) whose bandwidth parameter is adapted locally or across dimensions, often used in combinations to handle multiple feature scales.
Automatic Kernel Selection
Process integrated in MKL that identifies the most relevant subset of kernels from a candidate base, eliminating redundant or non-informative kernels through their learned weights.
Semantic Similarity Kernel
Specialized kernel type that encodes semantic relationships between entities (words, concepts), frequently combined with structural kernels in natural language processing applications.
Combined Gram Matrix
Final similarity matrix in a multiple kernel SVM, obtained by the weighted combination of individual Gram matrices of each kernel, serving as the basis for classifier optimization.
Diffusion Kernel
Graph theory-based kernel that captures diffusion similarity between nodes, often integrated in combinations to enrich representation with topological information.
Positive Semidefinite Optimization (SDP)
Class of convex optimization problems used to learn kernel weights under the constraint that the combined kernel matrix remains positive semidefinite, ensuring mathematical validity of the model.
Function-based Kernel
Approach where the combined kernel is defined as a weighted sum of predefined base functions, allowing clear interpretation of each similarity type's contribution to the final model.
Co-learning of Kernels
Strategy where multiple kernels are learned jointly by mutually training each other, each specializing on a subset of data to improve overall combination performance.
Multiple Kernel Cross-Validation
Evaluation method specific to MKL models where hyperparameter selection includes not only the SVM parameters but also the configuration and weights of the kernel combination.