Development of rule mining based classification models for quantitative data with many-objective intelligent metaheuristic optimization model

Classification rule mining, an important sub-field of data mining, aims to discover patterns in data sets that cannot be understood by the human perception. Today, most of the classical machine learning-based rule mining methods, despite their success, also have the handicap of being black-box models. Interpretability in machine learning methods, especially used in fields such as security, health and law, can be critical for decision makers. For this reason, studies on the development of interpretable or explainable artificial intelligence models have accelerated in recent years. Different techniques that add interpretability to machine learning methods continue to be proposed. However, among these proposed techniques, the use of metaheuristic optimization-based approaches has been limited. The advanced search mechanisms inherent in metaheuristic algorithms can offer flexible solutions. This thesis, seeing this gap in the literature, investigated the development of a new interpretable many-objective metaheuristic approach for classification rule mining. The most important originality of the thesis is the many-objective metaheuristic approach used for the first time in the field of rule mining. Unlike the approaches previously proposed, the developed approach can simultaneously optimize four different conflicting data mining metrics and provide interpretable solutions. The development of this method was carried out in stages during the thesis. Firstly, it is shown how metaheuristic approaches can be adapted to rule mining. These stages also provided the opportunity to test sub-mechanisms that will be used. In the last stage, the performance of the developed metaheuristic approach was tested comparatively with different machine learning methods. The results prove the success of this new rule mining approach.