Abstract: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.

Speed Up Kernel Projection Vector Machine Using Kronecker Decomposition