A new linear Support Vector Machine algorithm and solver are presented. The algorithm is in a twofold way well-suited for problems with a large number of training examples. First, unlike many optimization algorithms, it does not simultaneously keep all the examples in RAM and thus does not exhaust the memory (moreover, it smartly passes through disk files storing the data: two mechanisms reduce the computation time by disregarding some input data without a loss in solution quality). Second, it uses the analytical center cutting plane scheme, appearing as more efficient for hard parameter settings than the Kelley's scheme used in other solvers, like SVM_perf. The experiments with both real-life and artificial examples are described. In one of them the solver proved to be capable of solving a problem with one billion training examples. A critical analysis of the complexity of SVM_perf is given.
JavaScript jest wyłączony w Twojej przeglądarce internetowej. Włącz go, a następnie odśwież stronę, aby móc w pełni z niej korzystać.