Prototype-based classification by fuzzification of cases

Parisa KordJamshidi, Bernard De Baets, Guy De Tre.

A fuzzy prototype-based method is introduced for learning from examples. A special kind of prototype vector with fuzzy attributes is derived for each class from aggregating fuzzified cases for the purpose of concept description. The fuzzified cases are derived by defining a fuzzy membership function for each attribute of the sample cases. In a first method, for the classification of a new case, the membership degrees of its crisp attributes to fuzzy aggregated prototypes are measured. In a second method, after fuzzifying the new case, fuzzy set comparison methods are applied for measuring the similarity. The methods are compared to case-based ones like POSSIBL and kNN using UCI machine learning repository. We also make comparisons by using various transformation methods from probabilities to possibilities instead of defining membership functions.

PDF full paper