A comparison of the Web of Science and publication-level classification systems of science

- A classification system A is preferable to B when the normalization procedure based on the former performs better than the one based on the latter.
- Performance is assessed in terms of a graphical and a numerical test that use both classification systems for evaluation purposes.
- A publication-level algorithmically constructed system of 5,119 clusters is found to dominate a second one of 1,363 clusters.
- Although the system at the highest granularity level and the Web of Science journal-level system are non-comparable, we recommend the former.

In this paper, we propose a new criterion for choosing between a pair of classification systems of science that assign publications (or journals) to a set of clusters. Consider the standard target (cited-side) normalization procedure in which cluster mean citations are used as normalization factors. We recommend system A over system B whenever the standard normalization procedure based on system A performs better than the standard normalization procedure based on system B. Performance is assessed in terms of two double tests - one graphical, and one numerical - that use both classification systems for evaluation purposes. In addition, a pair of classification systems is compared using a third, independent classification system for evaluation purposes. We illustrate this strategy by comparing a Web of Science journal-level classification system, consisting of 236 journal subject categories, with two publication-level algorithmically constructed classification systems consisting of 1363 and 5119 clusters. There are two main findings. Firstly, the second publication-level system is found to dominate the first. Secondly, the publication-level system at the highest granularity level and the Web of Science journal-level system are found to be non-comparable. Nevertheless, we find reasons to recommend the publication-level option.