LATTICESVM - A new method for multi-class Support Vector machines

Zhibin Liu, Lianwen Jin

Abstract: Multi-class approaches for SVM (Support Vector Machines) is a very important issue for solving many practical problems (such as OCR and face recognition), since SVM was originally designed for binary class classification. Lots of methods based on traditional binary SVM have been proposed, each with its advantages and disadvantages. Among them, one-versus-one, one-versus-all, directed acyclic graph and binary tree are four most widely used methods. In this paper a novel LATTICESVM method, which can significantly reduce the storage and computational complexity, is proposed for multi-class SVM. A comparison in terms of storage, classification speed and accuracy against the four traditional multi-class approaches is given through both theoretic analysis and experiments on large scale handwritten Chinese character recognition. The results obtained clearly show the effectiveness of the proposed method.