Multi-class Classifier Based On Statistical Learning Research

Multiclass classification has always been an on-going issue in the machine learning community. With the increasing development of web and datamining technology, how to accurately and effectively solve multiclass problem with large number of classes has attracted attention to research and apply.Output coding is a general framework for solving multiclass problems. Most researches on output codes have focused on building multiclass machines using predefined output codes. One-against-all and One-against-one are most popular methods for multiclass problem. There were many researches on handling multiclass problem by solving one single optimization problem. Recently, some researchers have presented the notion of continuous codes and methods for designing output codes. They cast this problem as a constrained optimization problem, however these methods are time-consuming and expensive due to solving many optimization problems. In this paper, we present methods for multiclass problem that could overcome such limitation and solve the design problem effectively. We proposed a heuristic strategy using probability output of a minimax machine and margin information. We can view an algorithm with less iterative steps of optimization as a "weak" algorithm while preserving its other characteristic like large margin and geometric properties. Finally we make use of the kernel trick for "weak" algorithms' output (vector) to work in high dimensional spaces and improve the performances. We describe experiments with the proposed algorithm, comparing it with others. And an inspiring Experimental results show that our approach is competitive with others.