An Evolving Feature Weighting Framework for Granular Fuzzy Logic Models

Abstract

Discovering and extracting knowledge from large databases are key elements in granular computing (GrC). The knowledge extracted, in the form of information granules can be used to build rule-based systems such as Fuzzy Logic inference systems. Algorithms for iterative data granulation in the literature treat all variables equally and neglects the difference in variable importance, as a potential mechanism to influence the data clustering process. In this paper, an iterative data granulation algorithm with feature weighting called W-GrC is proposed. By hypothesising that the variables or features used during the data granulation process can have different importance to how data granulation evolves, the weight of each feature’s influence is estimated based on the information granules on a given instance; this is updated in each iteration. The feature weights are estimated based on the sum of within granule variances. The proposed method is validated through various UCI classification problems:- Iris, Wine and Glass datasets. Result shows that for certain range of feature weight parameter, the new algorithm outperforms the conventional iterative granulation in terms of classification accuracy. We also give attention to the interpretability-accuracy trade-off in Fuzzy Logic-based systems and we show that W-GrC produces higher classification performance - without significant deterioration in terms of its interpretability (Nauck’s index).