The performance of kernel methods based pattern analysis tasks depends on learning a suitable kernel function that captures the similarities between the data examples. Many methods have been proposed for learning a suitable kernel matrix for a given labeled training set. In this work we discuss a novel approach for learning a kernel gram matrix using distance metric learning (DML) techniques from a sparingly labeled training set. We formulate the semi-supervised DML (SS-DML) problem in which the label information is available for a limited number of training examples only. This iterative method uses the DML technique to learn a kernel gram matrix by using pair wise similar/dissimilar constraints deduced out of the available limited number of labeled training examples. The label information for all the unlabeled examples is incrementally determined using an iterative self training strategy that uses an SVM whose output values are converted into posterior probabilities. This SVM uses a confidence measure in the decision making process. The SS-DML method learns an optimal kernel gram matrix from the available sparingly labeled training data by claiming both the benefits of DML and the superiority of SVM as a classifier. The SS-DML method can be effectively applied in tasks that are used in pattern analysis applications. © 2017 IEEE.