COMPARISON OF ARTIFICIAL NEURAL NETWORKS AND TRADITIONAL CLUSTERING METHODS

Abstract

Classification of objects proceeding from their numerical characteristics is considered to be the main tool of modern qualitative chemical analysis. Classification is widely used to extract useful information from multivariate experimental data for foodstuff, drugs, environmental objects, materials, substances, industrial wastes, etc. Artificial neural networks have received much attention recently. Thanks to their adaptive structure and learning capability, they are success fully used to solve classification, identification and prediction tasks [1, 2]. A new clustering procedure based on the combination of the unsupervised Kohonen and probabilistic artificial neural networks The approach been demonstrated to be efficient for the classification of a large set of solvents. The additional use of the leave-one-out cross-validation procedure has improved the results. The final solvent classification is meaningful and chemically interpretable