LaRank is a multi-class support vector machine training algorithm for approximate online and batch learning based on sequential minimal optimization. For batch learning, LaRank performs one or more learning epochs over the training set. One epoch sequentially tests all currently excluded training examples for inclusion in the dual optimization problem, with intermittent reprocess optimization steps on examples currently included. Working set selection for one reprocess step chooses the most violating pair among variables corresponding to a random example. We propose a new working set selection scheme which exploits the gradient update necessarily following an optimization step. This makes it computationally more efficient. Among a set of candidate examples we pick the one yielding maximum gain between either of the classes being updated and a randomly chosen third class. Experiments demonstrate faster convergence on three of four benchmark datasets and no significant difference on the fourth.
展开▼
