On a modified graph-theoretic partitioning method of cluster analysis

A modified method of cluster analysis based on graph theory is presented. The original clustering technique developed by H. SPATH [16] starts with pair wise distances between objects. These distances are used as the valuation of the edges in the complete graph that represents the data set. The partition one is looking for is an optimised set of complete subgraphs. The optimum property is defined concerning the edge valuation of the subgraphs. The proposal presented here is a more general one. Additionally, the valuation of vertices will be taken into consideration. In this way, both the connection between objects and the objects themselves are valued. The method is applied on data of chemical compounds of Roman bricks and tiles. As this example demonstrates it can be supposed that the degree of independence on the choice of the start partition is greater if the classification starts using a complete graph with valuation of both edges and vertices.