A Novel Approach to Estimate Posterior Probabilities by Class Conditional Confidence Transformations
Zhen Lou · Chinese Journal of Computers · 2005
It is a crucial problem to estimate posterior probabilities for combination techniques in pattern recognition. This paper proposes a novel class conditional confidence transformation approach to estimate posterior probabilities for 1NN distance classifiers. In the paper, posterior probabilities are assumed to be distributed on the nearest neighbor class and the second nearest neighbor class. It is supposed that there exists a class conditional confidence transformation function for each class. This paper gives an estimation approach to the class conditional confidence transformation functions by using training samples. Experiments have been performed with Concordia University CENPARMI’s handwritten digit database and Nanjing University of Science and Technology’s handwritten digit database. Experimental results show that there is a great deal of reason in the proposed approach to estimate class conditional confidence transformation functions and the novel approach is superior to poll, count, the linear approach and the adaptive confidence transform (ACT) approach in decreasing the erroneous classification rates.