A proposed framework for incorporating ´weight of evidence´ within a fuzzy rule based classification system

Two different strategies for the firing and weighting of rules in a fuzzy rule based classification system (FRBCS) are examined. The rules are automatically obtained by the "pattern discovery" (PD) by A.K.C. Wong and Y. Wang (1997) algorithm and applied via input membership functions to a FRBCS by A. Hamilton-Wright and D. W. Stashuk (2005). Classification performance of the FRBCS when using an occurrence based rule weighting and standard rule firing logic was compared to performance obtained when using a weight of evidence (WOE) based rule weighting with an "independent" rule firing strategy. Performance was studied across various linearly and non-linearly separable class distributions. A new defuzzification and rule firing framework required to accommodate the WOE rule weighting is described. The occurrence based rule weighting and standard rule firing strategy performed better than WOE rule weighting and "independent" rule firing. Variance in the estimation of the WOE statistic, related to the number of training examples available, may limit the performance of the WOE scheme.