SVM - Neighbor based candidate working set selection applied on text-categorization

Computational complexity is one of the most important issues in any machine-learning algorithm. A novel working set selection mechanism is proposed to improve Support Vector Machine (SVM) learning. Implementation is based on the Keerthi et al.´s SMO algorithm, but our approach is one-class classification. When selecting samples for the optimization process, much effort is spent to find the most violating pair. The training time strongly depends on the selection of these variables. By choosing the neighbor samples of the updating pair (current working set) one can reach the optimal solution much faster. This one-class classification approach will be applied to a text categorization problem using the pointwise total correlation for term indexing.