REPOZITORIJ > REZULTATI

Povzetek

The goal of knowledge discovery in databases is to construct models or discover
interesting patterns in data. Model construction and pattern discovery
are frequently performed by rule learning, as the induced rules are easy to be
interpreted by human experts. The standard classification rule learning task is
to induce classification/prediction models from labeled examples.
In contrast to predictive rule induction, where the goal is to induce a model in
the form of a set of rules, the goal of descriptive rule induction is to discover
individual patterns in the data, described in the form of individual rules.
This thesis introduces the term supervised descriptive rule induction (SDRI), as
a unification of several areas of machine learning that deal with finding comprehensible
rules from class labeled data. We developed a unifying framework for
contrast set mining, emerging pattern mining and subgroup discovery, as representatives
of supervised descriptive rule induction approaches, which includes
the unification of the terminology, definitions and heuristics. By using our SDRI
framework, we overcame some open issues and limitations of SDRI sub-areas,
like presenting the results to end users by visualization and supporting factors.
Applications of SDRI methods to real-life datasets, which in collaboration with
domain experts led to new insights in the analyzed domains and to methodology
developments, are also demonstrated. A new method called mining of
closed sets for labeled data (with the algorithm RelSets) and its application in
microarray data analysis is also presented. The main algorithms used in the
experiments were made available on-line.