Optimizing feature extraction for multiclass problems

Feature extraction has been an important topic in pattern classification and studied extensively by many authors. Most conventional feature extraction methods are performed using a criterion function between two classes or a global function. Although these methods work relatively well in most cases, generally it is not optimal in any sense for multiclass problems. In this paper, we propose a method optimizing feature extraction for multiclass problems. We first investigated the distribution of the classification accuracy of multiclass problems in the feature space and found that there exist much better feature sets that conventional feature extraction algorithms fail to find. Then we propose an algorithm that finds such features. Experiments show that the proposed algorithm consistently provides a superior performance compared with the conventional feature extraction algorithms.