EVALUATING CLASSIFIER COMBINATION USING SIMULATED CLASSIFIERS
Date
2000-10-25
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The use of standard data sets with known properties is standard
practice for the evaluation of composite classifiers. However, because the
properties of the original test data cannot be specified in advance, it is
difficult to conduct controlled tests on classifier combination methods. In
these cases it is common to use simulated data, but a standard means for doing
this has not evolved. Confusion matrices can also be used to create simulated
classifiers having similar properties to the original and, by generating
specific confusion matrices and basing a set of simulated classifiers on these,
test suites can be designed that have pre-determined properties. This
simulation allows controlled testing of classifier combination techniques.
Such a scheme is described and evaluated using known combination methods and
standard data sets, both to confirm the simulation and to demonstrate its
flexibility.
Description
Keywords
Computer Science