Graph-based clustering based on cutting sets

Abstract

One of the most prominent challenges in data mining is the clustering of databases containing many categorical attributes. Representation of such data in continuous, Euclidean space usually does not reflect the true segments of data. As a crucial consequence, clustering algorithms working in continuous, Euclidean space may produce segmentations of poor quality. An alternative direction explores graph-based representation of data. In this paper, we show that graph-based data representation is well suitable for the case of categorical attributes. In particular, we offer the following contributions: i) we propose and analyze a flexible graph-based genetic clustering algorithm, where the ideal clusters can be characterized using external cluster quality functions, called kernels, ii) we study kernels, and define the crucial property of effective kernels, iii) we introduce a framework for distributed data-oriented graph clustering computations. In contrast of the complexity of our problem, which turns out to be NP-hard in our analysis, experiments show that in case of well clusterable data, our algorithm has attractive scalability properties. We also perform experiments on real medical data that provides us with further evidence about the practical applicability of our approach.