Optimality of decision boundary feature extraction for multiclass problems
Jinwook Go, Chulhee Lee · 2004
In the decision boundary feature extraction algorithm, it was shown that useful information for discriminating among classes is related to decision boundaries and the method was extended to multiclass problems by combining the decision boundary feature matrix of each pair of classes. However, it was reported that this method could result in performance loss since some decision boundaries between two classes might be irrelevant in a multiclass problem. In this paper, we propose a method to eliminate this inefficiency and examine whether such occurrence affects the performance of the decision boundary feature extraction algorithm in multiclass problems.