Alpha seeding for support vector machines
Dennis DeCoste, Kiri L. Wagstaff · 2000
A key pra ti al obsta le in applying support ve tor ma-hines to many large-s ale data mining tasks is that SVM training time generally s ales quadrati ally (or worse) in the number of examples or support ve tors. This omplexity is further ompounded when a spe i SVM training is but one of many, su h as in Leave-One-Out-Cross-Validation (LOOCV) for determining optimal SVM parameters or as in wrapper-based feature sele tion. In this paper we explore new te hniques for redu ing the amortized ost of ea h su h SVM training, by seeding su essive SVM trainings with the results of previous similar trainings.