A stopping criteria for the Growing Neural Gas based on a validity separation index for clusters
Diego Chavez, Guillermo Laures, Kristel Loayza, Raquel E. Patiño-Escarcina · 2011
Data clustering is a very known problem in the machine learning area, there are a lot of methods that are capable to achieve this task, one of them is the Growing Neural Gas network, which is an unsupervised incremental clustering algorithm, this network is able to learn the important topological relations in a given set of input vectors by means of a simple Hebb-like learning rule. In contrast to previous approaches, this model has parameters which changes over time and is able to continue learning, adding units and connections and occasional removal of units, until a performance criterion can tells it when to stop, without taking into account the iterations number. This article proposes a new stopping criteria for the Growing Neural Gas, based on the error rate of a novel validation index for clusters called SV index, which indicates how much compact are each cluster within each self, and how much separated are between them. The results of the experiments shows that the stopping criteria really avoid unnecessary training, pointing out the validation of the current proposal.