A new momentum minimization decomposition method for support vector machines

The support vector machine classifier is a binary classifier applied to classify large datasets, which is ideal for the application of decomposition methods when processing memory is limited. However, the rates of convergence of the decomposition method are largely dependent on the sequence of decomposed problems solved. Unfortunately, choosing the optimal sequence of sub problems is difficult due to the inability of the algorithm to consider the entire variable space at once. We propose a measure of iteration that we call momentum and derive a prediction method to minimize the momentum of the updated iterates hitting the boundary constraints. Our prediction method uses a rough heuristic set to choose an approximately optimal subproblem to solve. We show that this rough heuristic set could greatly improve the speed of the popular sequential minimal optimization algorithm.