CLUSTERING OF SI-29 NMR DATA OF ALUMINOSILICATE GLASSES USING KAPPA-MEANS AND FUZZY C-MEANS - A COMPARISON

The results of applying k-means (KM) and fuzzy c-means (FCM) methods of cluster analysis to a set of Si-29 NMR data of calcium aluminosilicate glasses are compared. Hard partitions are obtained as a result of KM cluster analysis, i.e. each object belongs to only one definite cluster. The FCM algorithm optimizes the degree of membership of objects in all the clusters - each object may thus belong to more than one cluster. The results of FCM are influenced by a parameter which controls the 'degree of fuzziness'. This influence is studied and a method is suggested which makes the manifold of data obtained by FCM convenient and clear. The discussion of the partitions obtained shows that the FCM analysis yields valuable additional information on outliers and hybrids, i.e. less typical members of the clusters. The results of KM and FCM cluster analysis support the interpretation of the Si-29 NMR data on the basis of a structural model for the calcium aluminosilicate glasses.