Learning Distance Metric for Support Vector Machine: A Multiple Kernel Learning Approach

Recent work in distance metric learning has significantly improved the performance in k-nearest neighbor classification. However, the learned metric with these methods cannot adapt to the support vector machines (SVM), which are amongst the most popular classification algorithms using distance metrics to compare samples. In order to investigate the possibility to develop a novel model for joint learning distance metric and kernel classifier, in this paper, we provide a new parameterization scheme for incorporating the squared Mahalanobis distance into the Gaussian RBF kernel, and formulate kernel learning into a generalized multiple kernel learning framework, gearing towards SVM classification. We demonstrate the effectiveness of the proposed algorithm on the UCI machine learning datasets of varying sizes and difficulties and two real-world datasets. Experimental results show that the proposed model achieves competitive classification accuracies and comparable execution time by using spectral projected gradient descent optimizer compared with state-of-the-art methods.