The effect of data standardization in cluster analysis

Abstract

The application of multivariate techniques to experimental results requires a responsibility on behalf of the researcher to understand, evaluate and interpret their results, especially the ones that are more complex. The objective of this article is to evaluate the impact of three standardization techniques on the formation of clusters by means of the Kohonen neural network were studied. The standardization techniques studied were logarithm (log), generalized-log and improved minimum-maximum. The studies were performed using two different databases consisting of 298, named B1, and 146 samples, named B2. The B1 dataset is formed by samples that form two cluster very close. However, the B2 dataset form three diferent and separated cluster. The mass fractions of As, Ce, Cr, Cs, Eu, Fe, Hf, K, La, Lu, Na, Nd, Sc, Sm, Tb, Th, U, and Yb of each sample were determined by instrumental neutron activation analysis, INAA. Three validation indices : Jaccard, Fowlkes-Mallows and Rand were performed on the dataset. The results suggest that when the cluster are close, the improved minimum-maximum satandardization is better than the logarithm and generalized-log. However, when the cluster are separated, the logarithm and generalized-log are better than the improved minimum-maximum technique.