CLUSTER ANALYSIS USED ON SOCIAL SCIENCE JOURNAL CITATIONS

Abstract

Cluster analysis was used on three files of citations from social science journals to other journals. The files were a pilot study, a file of criminology data and a very large file covering all social sciences. The criminology data was divided into sections drawn from 1950, 1960, and 1970 sources. The large file was in two sections, one drawn from a ranked list of source journals and the other from a list of journals selected at random. The study looked at the effect of several cluster methods and various ways of normalizing the data to find out which observed effects are true properties of the data. The results indicated that clusters of social science journals generated using citations have a non‐hierarchical structure. The criminology samples from 1960 and 1970 showed little change over ten years in the main clusters, but the two sections of the large file gave results which, although similar in general shape, differed substantially in their details. The overall conclusion is that cluster analysis is an unsuitable approach to the design of secondary services in the social sciences, though it may have some value in automatic retrieval systems. Two problems are the vast amounts of data needed and the difficulty of presenting results comprehensibly, particularly with overlapping clusters.