Fusion of neural gas

Abstract

One of the most important feature of the Neural Gas is its ability to preserve the topology in the projection of highly dimensional input spaces to lower dimensions vector quantizations. For this reason, the Neural Gas has proven to be a valuable tool in data mining applications. In this paper an incremental ensemble method for the combination of various Neural Gas models is proposed. Several models are trained with bootstrap samples of the data, the "codebooks" with similar Voronoi polygons are merged in one fused node and neighborhood relations are established by linking similar fused nodes. The aim of combining the Neural Gas is to improve the quality and robustness of the topological representation of the single model. We have called this model Fusion-NG. Computational experiments show that the Fusion-NG model effectively preserves the topology of the input space and improves the representation of the single Neural Gas model. Furthermore, the Fusion-NG explicitly shows the neighborhood relations of it prototypes. We report the performance results using synthetic and real datasets, the latter obtained from a benchmark site. © Springer-Verlag Berlin Heidelberg 2007.