Alpha seeding for support vector machines

A key pra obsta in applying support ve ma-hines to many large-s data mining tasks is that SVM training time generally s quadrati (or worse) in the number of examples or support ve. This is further ompounded when a spe SVM training is but one of many, su as in Leave-One-Out-Cross-Validation (LOOCV) for determining optimal SVM parameters or as in wrapper-based feature sele. In this paper we explore new te for redu the amortized of ea su SVM training, by seeding su SVM trainings with the results of previous similar trainings.