A procedure for meaningful unsupervised clustering and its application for solvent classification

Abstract

AbstractArtificial neural networks have proven to be a powerful tool for solving classification problems. Some difficulties still need to be overcome for their successful application to chemical data. The use of supervised neural networks implies the initial distribution of patterns between the pre-determined classes, while attribution of objects to the classes may be uncertain. Unsupervised neural networks are free from this problem, but do not always reveal the real structure of data. Classification algorithms which do not require a priori information about the distribution of patterns between the pre-determined classes and provide meaningful results are of special interest. This paper presents an approach based on the combination of Kohonen and probabilistic networks which enables the determination of the number of classes and the reliable classification of objects. This is illustrated for a set of 76 solvents based on nine characteristics. The resulting classification is chemically interpretable. The approach proved to be also applicable in a different field, namely in examining the solubility of C60 fullerene. The solvents belonging to the same group demonstrate similar abilities to dissolve C60. This makes it possible to estimate the solubility of fullerenes in solvents for which there are no experimental data