SVM vs LS-SVM: chi vince?
A questa frequente domanda di chi approccia al mondo delle SVM, riporta uno stralcio di risposta di uno degli autori delle LS-SVM (Suykens):
What’s the advantage and disadvantage of LS-SVM compared with standard SVM ?
One of the main motivations of LS-SVM approaches is to make SVM methodologies more generally applicable (trying to simplify in order to be able to extend) in a similar spirit as classical neural networks (such as MLPs and RBF networks which can be used in classifiers, feedforward and recurrent nets, unsupervised learning, control etc.)
Primal-dual LS-SVM formulations have been given e.g. to classifiers (related to kernel FDA), to function estimation (equivalent to RN, GP, RKHS, KRR), weighted versions for robust estimation, Bayesian inference and probabilistic interpretations, kernel PCA,PLS,CCA, recurrent networks, optimal control. A version which is very suitable for on-line and fast adaptive signal processing, large scale problems and transductive inference is `fixed-size LS-SVMs’ (which are sparse approximation models like standard SVMs) and make use of Nystrom approximation (as known in the GP area).
In this way we aim at creating a unifying framework and interdisciplinary avenue of primal-dual modelling thinking in relation to areas as statitics, signal processing, datamining, systems and control, signal processing, machine learning, pattern recognition, mathematics and many other application areas. In other words trying to get the big picture…
