Partially supervised learning of fuzzy classification rules

The research area of Data Mining or Knowledge Discovery in Databases has emerged in response to the challenges of analyzing the tremendously growing datasets gathered nowadays by companies and research institutions. Classification is one important task of data mining, where fuzzy techniques to extract classification rules from data are appealing due to their human understandable modeling. Often, datasets to be analyzed do not contain class labels, and their size renders manual labeling infeasible. Thus, there is an increasing interest in semi-supervised methods that can learn from only partially labeled data. Unfortunately, most current data mining methods are supervised, and most current semi-supervised methods do not generate human understandable models. In this thesis we review the key concepts of fuzzy classification and fuzzy classifier learning, with a focus on their capabilities and interpretability, and reveal some common peculiarities and pitfalls. Furthermore, we review approaches to semi-supervised learning with a stress on fuzzy methods, and show their deficiencies, particularly for inducing interpretable fuzzy rules. The main achievements of this thesis are the development of an evolutionary algorithm and specialized fitness functions that allow semisupervised learning of interpretable fuzzy rules.