Reducing large diagonals in kernel matrices through semidefinite programming

Hsiao-Mei Lu, Sumeet Gupta, Yang Dai

Abstract: Classification problems in bioinformatics solved by Support Vector Machine (SVM) learning algorithms may result in kernel matrices with very large diagonal elements. This structural characteristic of the kernel matrix leads to overfitting. A method for reducing large diagonals based on the use of Semi-Definite Programming (SDP) is proposed. The kernel matrix generated by SDP is guaranteed to be positive semidefinite. The preliminary experimental performance of the new kernel is encouraging.