Sifter: an approach for robust fuzzy rule set discovery

One of the problems in the field of data mining with evolutionary algorithms is the variance of accuracy in multiple runs. Decreasing the variance of accuracy without any accuracy reduction is very difficult since there is a trade-off between these conflicting objectives. In this paper we follow two abstract objectives: accuracy and interpretability. The interpretability is measured by three criteria: number of the rules, sum of the rules lengths and the standard deviation of the accuracy (Acc.SD). The proposed method consists of two stages, and in both, an innovative binary version of the krill herd algorithm has been introduced. In this study, choosing the best krill in population and indicating the local best of the krills in each generation are performed according to a new multi-objective function. In the first stage, candidate rules are generated intelligently using Pittsburgh and iterative rule learning approaches that guarantee the diversity of the extracted rules. The Sifter approach, that is presented here, uses a clustering concept and is incorporated in stage two for robust rule set selection from the candidate rules. Multiple executions of Sifter give roughly the same results. Also in this study, we offer the rule set distance measure that is calculated in two modes: Morphologically and Semantically. Experimental results show that we have successfully improved the two objectives that are naturally in conflict.