Support Vector Machines for Regression Problems with Sequential Minimal Optimization

Training a support vector machine (SVM) is usually done by mapping the underlying optimization problem into a quadratic programming (QP) problem. Unfortunately, high quality QP solvers are not readily available, which makes research into the area of SVMs difficult for the those without a QP solver. Recently, the Sequential Minimal Optimization algorithm (SMO) was introduced [1, 2]. SMO reduces SVM training down to a series of smaller QP subproblems that have an analytical solution and, therefore, does not require a general QP solver. SMO has been shown to be very efficient for classification problems using linear SVMs and/or sparse data sets. This work shows how SMO can be generalized to handle regression problems.