Single Pass SVM using Minimum Enclosing Ball of Streaming Data

We present a stream algorithm for large scale classification (in the context of l2-SVM) by leveraging connections between learning and computational geometry. The stream model [1] imposes the constraint that only a single pass over the data is allowed. We study the streaming model for the problem of binary classification with SVMs and propose a single pass SVM algorithm based on the minimum enclosing ball (MEB) of streaming data [2]. We show that the MEB updates for the streaming case can be adapted to learn the SVM weight vector using simple Perceptron-like update equations. Our algorithm performs polylogarithmic computation at each example, requires very small and constant storage (O(D) where D is the dimensionality of input space). Experimental results show that, even in such restrictive settings, we can learn efficiently in just one pass and get accuracies comparable to other state-of-the-art SVM solvers. The 2-class l2-SVM [3] is defined by a hypothesis f(x) = w φ(x), and a training set consisting of N points {zn = (xn, yn)}n=1 with yn ∈ {−1, 1} and xn ∈ R. The only difference between the l2-SVM and the standard SVM is that the penalty term has the form (C ∑