Exploring the space of coding matrix classifiers for hierarchical multiclass text categorization
One of the ways of approaching a multiclass classification problem is to transform it into several two-class (binary) classification problems. An ensemble of binary classifiers is trained for these tasks and their predictions are combined using a voting method into predictions for the original multiclass problem. Each of the new binary problems uses some of the original classes as positive training data, some classes as negative training data and the remaining classes (if any) are not used at all. The relationship between classes (of the original problem) and binary classifiers can be concisely represented by a matrix called the coding matrix. In this paper we explore some of the statistical properties of the space of coding matrix based classifiers in the context of a small hierarchical multiclass learning problem.