Semi-supervised Metric Learning Using Composite Kernel

Learning an appropriate distance metric using the available class labels or some other supervisory information is a very active research area. It has been shown that the metric learning based methods outperforms the traditionally used distance metrics such as the Euclidean distance metric. In kernelized version of metric learning algorithms, the data is implicitly transferred into a new feature space using a non-linear kernel function. The distance metric learning process is performed in the new feature space. Selecting an appropriate kernel function and/or tuning its parameters impose significant challenges in the kernel-based methods. Toward this goal, we present a semi-supervised metric learning algorithm using composite kernels. We demonstrate the usefulness of the proposed method on both synthetic and real-world data sets.