Interactive genetic fuzzy rule selection through evolutionary multiobjective optimization with user preference

The design of fuzzy rule-based classifiers from numerical data can be regarded as one of data mining approaches. The main characteristic is its knowledge representation. Antecedent fuzzy sets of each fuzzy if-then rule linguistically represent a fuzzy region in the pattern space. When a user´s main concern is knowledge extraction rather than classifier design, we have to consider two conflicting objectives: accuracy maximization and interpretability maximization. The number of correctly classified training patterns is often used for accuracy measure. On the other hand, interpretability is very subjective and hardly defined without the user´s involvement. In this paper, we incorporate user´s preference information into multiobjective genetic fuzzy rule selection. We propose a preference function composed of satisfaction level functions on six criteria: average confidence, average coverage, the number of used attributes, the maximum number of used granularities, classification accuracy, and the number of rules. Since it is hard to directly define each satisfaction level function beforehand, it is interactively modified by the user during the evolution. The preference function is handled as one of objective functions in multiobjective genetic fuzzy rule selection. The effectiveness of the proposed method is examined through a case study for knowledge extraction from the Pima Indian Diabetes data.