EXPECTED KERNEL FOR MISSING FEATURES IN SUPPORT VECTOR MACHINES

The expected kernel for missing features is introduced and applied to training a support vector machine. The expected kernel is a measure of the mean similarity with respect to the distribution of the missing features. We compare the expected kernel SVM with the robust second-order cone program (SOCP) SVM, which accounts for missing kernel values by estimating the mean and covariance of missing similarities. Further, we extend the SOCP SVM to utilize the expected kernel by deriving the expected kernel variance. Results show that the expected kernel-used with a traditional SVM solver-shows competitive performance on benchmark datasets to the SOCP SVM at a far-reduced computational burden.