Abstract
In clusters analysis, a problem of great interest is having methods that allow the representation of the topology of input space without the need to know additional information about it. This gives rise to growing competitive neural methods which are capable of determining the structure of the network autonomously during the process of training. This work proposes a variation of the Growing Neural Gas (GNG) algorithm, calling GNG with post-pruning (GNG-PP), and a method of clustering based on the search for topological neighborhoods generated by the former. These were combined in a three-phase process to clustering the S&P100 set, which belongs to the macroeconomic field. This problem has a high dimensionality in the characteristics space. Its results are compared to those obtained by SOM, Growing Cell Structures (GCS), and a non-neural method. Evaluation of the results was made by means of the kappa coefficient, using as evaluation set the GICS industrial classification. The results show that when using the proposed methods the best clustering are generated, obtaining a kappa coefficient of 0.5643 in the GICS classification. © Springer-Verlag Berlin Heidelberg 2007.
