In this thesis, we show a classification procedure using methodologies of combinatorial optimization to partition the euclidean space into convex sets of prescribed number and size. After introducing the theoretical background, we present a clustering based classifier and compare it with established algorithms. We show that an iterative sequence based on a geometric clustering in each step leads to a segmentation of the data space especially suitable for our prediction task. Based on this procedure, we define corresponding binary classifiers and introduce a new probabilistic test procedure to evaluate the reliability of a clustering based prediction. Furthermore, we show the excellent performance of the new classification technique and demonstrate the clustering based test of hypotheses on real world data.
«In this thesis, we show a classification procedure using methodologies of combinatorial optimization to partition the euclidean space into convex sets of prescribed number and size. After introducing the theoretical background, we present a clustering based classifier and compare it with established algorithms. We show that an iterative sequence based on a geometric clustering in each step leads to a segmentation of the data space especially suitable for our prediction task. Based on this proced...
»