Speed Up Kernel Projection Vector Machine Using Kronecker Decomposition

摘要

We present a speedup algorithm for kernel projection vector machine (KPVM) based on kronecker product decomposition. The large scale kernel matrix K with size of n × n is factorized into two small matrices K1 and K2 with size n1 × n1 and n2 × n2 respectively where n1 × n2 = n. The time-consuming SVD operation on K in KPVM is calculated through K1 and K2. The computation complexity is reduced to O(n2} from O(n3) originally while generalization ability is undiminished or even better than KPVM.