svm-subgradient-large-n-coorddescent-high-dim

IN premise — entries/2026/06/21/wiki-Support_vector_machine-chunk-3.md

Created 2026-06-21T09:55:54+00:00

Sub-gradient descent is preferred for SVMs with large n (many training examples); coordinate descent is preferred for high-dimensional feature spaces.

Dependents

These beliefs depend on this one:

• IN svm-solver-ecosystem-scale-dependent — SVMs have a rich ecosystem of specialized solvers whose optimal choice depends on problem geometry — SMO decomposes into 2D subproblems for general QP, Pegasos uses stochastic gradient descent in the primal for large-scale problems, LIBLINEAR achieves linear-time iterations for high-dimensional sparse data, and the crossover point between sub-gradient and coordinate descent depends on whether n or d dominates.