Abstract:
Classification problems are naturally characterized by uncertainty, subjectivity, imprecision
and ambiguity. Thus, in designing classification models, mathematical methods
that are able to satisfactorily deal with uncertainty, ambiguity and subjectivity are
essential. Although fuzzy set theory is a very convenient mathematical tool for treating
vagueness and ambiguity due to redundant and irrelevant features, and poor class
definition, the existing fuzzy entropy based techniques can not effectively deal with
ambiguity. This thesis presents a geometrical fuzzy similarity classifier which allows
us not only reduce complexity of classification problems by removing redundant and
irrelevant features, but also estimate ambiguity in class assignment using measures of
fuzzy specificity. The model was tested using 4 benchmark datasets from University
of California Irvine (UCI) machine learning repository yielding very attractive results.
With Dermatology data set, a mean classification accuracy of 98:21% was obtained
with only 24 features as compared with 97:82% with 34 features. Uncertainty associated
with class assignment is also reported. Miss-classified samples display high
average uncertainty as compared to those correctly classified.
