A Parallel Algorithm for SVM Based on Extended Saddle Point Condition

Xiaorui Li, Congying Han, Guoping He

Abstract: The constraint-partitioning approach achieves a significant reduction in solution time while resolving some large-scale mixed-integer optimization problems. Its theoretical foundation, extended saddle-point theory, which implies the original problem can be decomposed into several subproblems of relatively smaller scale in virtue of the separability of extended saddle-point conditions, still needs to be deliberated carefully. Enlightened by such a plausible theory, we have developed a novel parallel algorithm for convex programming. Our approach not only works well theoretically, but also may be promising in numerical experiments. As the theoretical essence of Support Vector Machine (SVM) is a quadratic programming, we are inspired to apply this new method onto large-scale SVMs to achieve some numerical improvements.