A three-way classification strategy for reducing class-abundance: the zip code recognition example
Chun‐Houh Chen, Ker-Chau Li · Lecture notes-monograph series · 2004
For many real world problems in discriminant analysis, the abundance of classes is one major source of complexity.Let k be the number of classes.When k is large, it is difficult to discriminate all k classes at a time.To ignite the statistical community's interest in addressing this issue, we focus on the handwritten digit recognition problem using the Zip code data analyzed in the seminal article by LeCun et al (1989) wherein the neuron network with backpropagation was first applied successfully to a real world problem.We propose a three-way sub-classification strategy that first divides the original A -class problem into C(k, 2) = k(k -l)/2 smaller problems.Each smaller problem concerns the discrimination between two classes from the initial k classes under the presence of a third class that consists of all cases from the rest of the k -2 classes.The decisions from each sub-problem of three-way classification are then combined via a conditional error rate analysis that exploits the degree of consensus among various decisions.We report the existence of a large portion of high quality images found by our method; they are predicted with an error rate less than 1%.