Stanford University - Department of Statistics
This talk is in two parts. I n the first I will explore a new classification technique called the Error Correcting Output Coding method (ECOC). As the name suggests the original motivation for this procedure relyed on Error Coding ideas. I will show results which suggest that this is not the reason for its success. In the process I will introduce a new classifier called the Substitution method which works on a similar principle to ECOC but generally outperforms it. In the second part I will talk about Majority Vote Classifiers (MVC). I show that the Substitution method is an example of an MVC and that the ECOC method is an "approximate" MVC. I will also detail some of the theories that have been suggested to explain the success of Majority Vote Classifiers.