Handling High Dimensionality and Interpretability-Accuracy Trade-Off Issues in Evolutionary Multiobjective Fuzzy Classifiers

Fuzzy systems are capable to model the inherent uncertainties in real world problems and implement human decision making. In this paper two issues related to fuzzy systems development are addressed and solutions are proposed and implemented. First issue is related to the high dimensional data sets. Such kinds of data sets lead to explode the search space of generated rules and results into deterioration of interpretability and performance of the fuzzy classifiers. To deal with this problem several data clustering algorithms are developed, i.e. fuzzy c-mean clustering algorithm, entropy based fuzzy clustering algorithm etc. The authors have proposed an integrated version of clustering algorithm by replacing the cluster center generation method of fuzzy c-means algorithm by entropy based method. MATLAB is used to implement the proposed fuzzy clustering. On the other hand, interpretability is the subjective feature of fuzzy system that quantifies its understandability. Interpretability can be improved at the cost of other, i.e. improvement in one leads to loss in other. This situation is called ‘interpretability-accuracy trade-off’. By handling this trade-off a number of fuzzy systems can be generated with different values of interpretability and accuracy parameters. Evolutionary Multiobjective Optimization is used to deal with this trade-off by proposing a new fuzzy classifier ‘Teacher-Performance Fuzzy Classification System’. To implement fuzzy classifier ‘Guaje’ open access software is used and evolutionary multiobjective optimization framework is implemented using ‘MATLAB’.