Item

Abstract

We describe methods for taking into account unlabeled data in the training of a kernel-based classifier, such as a Support Vector
Machines (SVM). We propose two approaches utilizing unlabeled points
in the vicinity of labeled ones. Both of the approaches effectively
modify the metric of the pattern space, either by using non-spherical
Gaussian density estimates which are determined using EM, or by
modifying the kernel function using displacement vectors computed from
pairs of unlabeled and labeled points. The latter is linked to
techniques for training invariant SVMs. We present experimental
results indicating that the proposed technique can lead to substantial
improvements of classification accuracy.