Utvärdering av klusteranalytiska metoder i kombination med bibliografisk koppling
en bibliometrisk kartläggning av aktuella forskningsteman inom informationsvetenskap
The purpose of this thesis is to evaluate the combination of bibliographic coupling and three hierarchical cluster analytic techniques: average, complete and single linkage. Bibliographic coupling is applied as a measure of similarity between recently published articles and the cluster techniques are used for the partition of a set of articles into subsets of articles with similar research focus. The field of information science is applied as the test arena and the selection of data, i.e., the article population to be partitioned, is based on a number of central journals. Methods of evaluation include both quantitative approaches as well as a qualitative one. For comparing the resulting partitions, adjusted Rand index is utilized and for the establishment of cluster coherence and isolation, measures of average normalized coupling strength is applied. In an effort to quantify the degree of subject focus within clusters, normalized entropy of a cluster?s distribution of weighted descriptors is calculated. Manual inspection of cluster composition is also performed. The findings show that choice of cluster method has a profound impact on the resulting partitions. Complete linkage generated highly subject coherent clusters but produced a fragmented picture of the research field in question, i.e., the split up of research specialties. Single linkage on the other hand, performed unsatisfactory and is not recommended. It is suggested that average linkage, when combined with a stopping rule (the inconsistency coefficient), should be considered as the method of choice since it mainly generated reasonable subject coherent clusters while avoiding severe fragmentation.