Separable approximate optimization of support vector machines for distributed sensing

Abstract

Sensor measurements from diverse locations connected with possibly low bandwidth communication channels pose a challenge of resource-restricted distributed data analyses. In such settings it would be desirable to perform learning in each location as much as possible, without transferring all data to a central node. Applying the support vector machines (SVMs) with nonlinear kernels becomes nontrivial, however. In this paper, we present an efficient optimization scheme for training SVMs over such sensor networks. Our framework performs optimization independently in each node, using only the local features stored in the respective node. We make use of multiple local kernels and explicit approximations to the feature mappings induced by them. Together they allow us constructing a separable surrogate objective that provides an upper bound of the primal SVM objective. A central coordination is also designed to adjust the weights among local kernels for improved prediction, while minimizing communication cost. © 2012 Springer-Verlag.