Discriminative dictionary learning aims to learn a dictionary from training samples to enhance the discriminative capability of their coding vectors. Several discrimination terms have been proposed by assessing the prediction loss (e.g., logistic regression) or class separation criterion (e.g., Fisher discrimination criterion) on the coding vectors. In this paper, we provide a new insight on discriminative dictionary learning. Specifically, we formulate the discrimination term as the weighted summation of the squared distances between all pairs of coding vectors. The discrimination term in the state-of-the-art Fisher discrimination dictionary learning (FDDL) method can be explained as a special case of our model, where the weights are simply determined by the numbers of samples of each class. We then propose a parameterization method to adaptively determine the weight of each coding vector pair, which leads to a support vector guided dictionary learning (SVGDL) model. Compared with FDDL, SVGDL can adaptively assign different weights to different pairs of coding vectors. More importantly, SVGDL automatically selects only a few critical pairs to assign non-zero weights, resulting in better generalization ability for pattern recognition tasks. The experimental results on a series of benchmark databases show that SVGDL outperforms many state-of-the-art discriminative dictionary learning methods. ; Department of Computing