57 Kernel Logistic Regression Using Truncated Newton Method

Accurate prediction is important in data classification. The combination of penalized kernel logistic regression (KLR), truncated-regularized Newton method, iteratively re-weighted least-squares (TR-IRLS) has led to a powerful binary classification method using small-to-medium size datasets. Compared to support vector machines (SVM) and TR-IRLS on six benchmark publicly available datasets, the proposed algorithm is as accurate as, and much faster than, SVM, as well as more accurate than TR-IRLS. The algorithm also has the advantage of providing direct prediction probabilities.