An interpretable fuzzy-ensemble method for classification and data analysis / Adel Lahsasna

Despite the advantage of being highly accurate classifiers, many machine learning methods such as artificial neural networks (ANNs) and support vector machines (SVMs) have been criticized for their lack of interpretability as users are prevented from knowing about the decision process of their inner systems. Interpretability, which refers to the ability of a system to express its behavior in an understandable way, has recently gained more attention and it is considered as an important requirement especially for those applications that use knowledge-based systems such as decision support systems. The main objective of our study is to propose an interpretable fuzzy-ensemble method that can be used for both classification and data analysis. This classifier is the result of combining the advantages of an interpretable fuzzy rule-based system and accurate ensemble method. To achieve the aforementioned objective, we firstly propose two variant methods of a well-known fuzzy classifier proposed in (Ishibuchi & Nojima, 2007) aiming to improve its ability to maximize the accuracy of the fuzzy rule-based system while preserve its interpretability. In addition, we proposed a feature selection-based method that aims to improve the quality of the non-dominated fuzzy rule-based systems especially those generated from high dimensional data sets by allowing the genetic algorithm (GA) to start from a good initial population. For the ensemble method, we propose a design that combines five different base classifiers and use a GA-based selection method to select a subset from all the ensemble outputs using accuracy and diversity measures as two objectives in the fitness function. In addition, we propose a combination method that aims to improve the accuracy of the fuzzy rule-based system by using the accurate ensemble method to classify the patterns that have low certainty degree or in cases of rejected and uncovered classifications. iv The proposed method is tested using six data sets from the UCI machine learning repository, and the obtained results are compared with other benchmark methods. The results show that the fuzzy-ensemble method was able to maintain to a great extent the superiority of the ensemble method accuracy over the fuzzy rule-based system by successfully retaining an average of 76.77% of the accuracy gains obtained by the ensemble method relative to fuzzy rule-based system. In addition, the fuzzy-ensemble method has successfully preserved its interpretability compared to the fuzzy rule-based system. In addition, the two developed methods, namely, the fuzzy rule-based system and the ensemble method have shown separately competitive results with their related methods proposed in the literature. Thus, in addition to the proposed fuzzy-ensemble method, they can be separately used as single classifiers.