The nearest neighbor and the perceptron algorithms are intuitively motivated by the aims to exploit the "cluster" and "linear separation" structure of the data to be classified, respectively. We develop a new online perceptron-like algorithm, Pounce, to exploit both types of structure. We refine the usual margin-based analysis of a perceptron-like algorithm to now additionally reflect the cluster-structure of the input space. We apply our methods to study the problem of predicting the labeling of a graph. We find that when both the quantity and extent of the clusters are small we may improve arbitrarily over a purely margin-based analysis.