Support Vector Machines and Support Vector Regression › Disadvantages

Exact nonlinear kernel SVMs can become expensive as the training set grows. Materializing the full Gram matrix requires $O(n^2)$ storage, while decomposition solvers can compute entries on demand and keep a bounded cache. Training time depends on the kernel, feature dimension, data, cache, solver, and stopping tolerance; a single quadratic or cubic rate does not describe every SVM. Large linear problems can instead be fitted without a Gram matrix. Feature scales also affect the regularization and kernel geometry, so choose a meaningful scale for the inputs; standardization is a useful default when units differ, as discussed in Standardization.

SVM decision scores are not calibrated probabilities. In a binary model the score is $f(\mathbf{x})$; with a nonzero normal vector, divide by its feature-space norm to obtain signed distance to the hyperplane. Procedures such as Platt scaling or isotonic regression can estimate probabilities from scores, adding estimation work whose quality must be checked. Kernel settings and the penalty $C$ affect the fit, so compare relevant choices through hyperparameter search. With imbalanced classes, unweighted summed loss gives more total weight to the more numerous class; class weights or resampling within training folds can change that tradeoff.