We focus on an ensemble of graphical and statistical tools which represent the state of the art to assess the reliability of Self Organizing Maps. In particular, we are interested in methods that are able to provide information about: (a) the confidence we can give to the results of Self Organizing Maps; (b) the speed of convergence, depending on the existence of defined clusters within the data sample; and (c) conversely to (b), the possibility to infer the existence and significance of clusters from convergence behavior. We have found that some of the answers can be provided by three different techniques, namely, the STAB index suggested by Cottrell et al., the U-Matrix method of Ultsch and Vetter, and the CI index, introduced by the authors of this note. We will then try to evaluate the potential of those different methods, showing their points of contact (if any), as well as their major strengths or weaknesses. To such purpose, we will run simulations on various data samples, and discuss their results
Reliability and convergence on Kohonen maps: an empirical study
RESTA, MARINA;
2004-01-01
Abstract
We focus on an ensemble of graphical and statistical tools which represent the state of the art to assess the reliability of Self Organizing Maps. In particular, we are interested in methods that are able to provide information about: (a) the confidence we can give to the results of Self Organizing Maps; (b) the speed of convergence, depending on the existence of defined clusters within the data sample; and (c) conversely to (b), the possibility to infer the existence and significance of clusters from convergence behavior. We have found that some of the answers can be provided by three different techniques, namely, the STAB index suggested by Cottrell et al., the U-Matrix method of Ultsch and Vetter, and the CI index, introduced by the authors of this note. We will then try to evaluate the potential of those different methods, showing their points of contact (if any), as well as their major strengths or weaknesses. To such purpose, we will run simulations on various data samples, and discuss their resultsI documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.