An Improved Boundary Uncertainty-Based Estimation for Classifier Evaluation

Abstract This paper proposes a new boundary uncertainty-based estimation method that has significantly higher accuracy, scalability, and applicability than our previously proposed uncertainty method. In previous work, we introduced classifier evaluation metric termed “boundary uncertainty.” The name uncertainty” comes from evaluating the based solely on measuring equality between class posterior probabilities along boundary; satisfaction of such can be described as “uncertainty” boundary. We also to estimate this metric. By focusing evaluate its uncertainty, defines an easier target accurately estimated directly finite training set without using validation set. Regardless dataset, is defined 0 1, where 1 indicates whether probability for Bayes error achieved. call “Proposal 1” in order contrast it with paper, which 2.” Using Proposal performed successful real-world data supported theoretical analysis. However, suffered limitations owing difficulty finding location multidimensional sample space. novelty 2 locally reformalizes single dimension focuses convenient reduction focus toward provides method’s significant improvements. experiments Support Vector Machines (SVM) MultiLayer Perceptron (MLP), demonstrate offers competitive accuracy compared benchmark Cross Validation (CV) well much scalability both CV 1.