The comparative ability of self-organizing neural networks to define cluster structure

The ability to determine clusters or similarity in large, multivariate data sets is critical to many business decisions. Unfortunately, current cluster algorithms are sensitive to dispersion that occurs naturally in empirical data. As the level of relative cluster dispersion in the data increases, current clustering techniques fail to accurately identify cluster membership. An improved clustering methodology is needed that produces more accurate cluster definitions than the methods commonly used today. The research presented here investigates the ability of specific neural network architectures utilizing unsupervised learning to recover cluster structure from multivariate data sets with various levels of relative cluster dispersion. The results demonstrate that the Self Organizing Map network is a superior clustering technique and that its relative advantage over conventional techniques increases with higher levels of relative cluster dispersion in the data.