Mark Last

In this paper we introduce a method of creating structural (i.e. graph-based) classifier ensembles through random node selection. Different k-Nearest Neighbor classifiers, based on a graph distance measure, are created automatically by randomly removing nodes in each prototype graph, similar to random feature subset selection for creating ensembles of statistical classifiers. These classifiers are then combined using a Borda ranking scheme to form a multiple classifier system. We examine the performance of this method when classifying a web document collection; experimental results show the proposed method can outperform a single classifier approach (using either a graph-based or vector-based representation).