Hierarchical classification and granular regression using taxonomically structured background knowledge

This dissertation explores the power of knowledge-guided rule learning using taxonomic background knowledge in the form of concept-class taxonomies. It is shown that such taxonomic background knowledge can be effectively integrated into the rule learning process and that it does improve the predictive accuracy of the resulting classification models. Novel hierarchical rule-learning techniques are introduced that can be used to incorporate available background knowledge in the form of concept-class taxonomies into the learning process. The resulting models, known as multi-level hierarchical classification models, consist of both rules that make higher-level class membership predictions and rules that make lower-level class membership predictions. Empirical evidence is presented to show that multi-level hierarchical classification models, generated using concept-class taxonomies, have significantly higher predictive accuracies, when compared with models generated without using background knowledge, especially while dealing with small data sets with relatively large sets of concept-class values. However, it is argued that the improvements in predictive accuracies are obtained with corresponding tradeoffs in the granularity of the predictions.; A natural extension to the problem of hierarchical learning using taxonomically ordered concept-class values is the learning task involving continuous (numeric) concept-class values, typically handled by traditional regression methods. In many data-mining tasks, the end goal is not necessarily the generation of a regression model that predicts the actual numeric value of an unlabelled instance. Other relevant goals are detecting and identifying interesting ordered chunks of the independent variable, and generating comprehensible models mapping attribute values to the concept-class segments. This dissertation also explores the problem of rule learning using numeric concept-class values and discusses the granular regression rule learning problem and its related issues. A novel hierarchical granular regression rule-learning (HGRRL ) technique that addresses some of these limitations, is introduced.