Generating clustered journal maps: an automated system for hierarchical classification

Scholarly journals have been and remain the primary organizers of scientific communication. The number of journals has increased over the centuries, at times showing exponential growth (Mabe and Amin 2001; Mabe 2003; de Solla Price 1961, p. 166), but the journal form has remained remarkably stable in the social life of science. The intellectual development of the sciences and their organization, as well as growth of new specialties and disciplines, is organized, validated, and retained in scholarly journals. Ware and Mabe (2015) estimated that there were 28,100 peer-reviewed journals published in English in 2015; in the Web of Science (WoS) in that year, 11,365 journals were indexed. The source journals can be expected to account for more than 90% of citations because of the skew in the distributions (Garfield 1971; Seglen 1992).

Specialties and new fields develop at a level above individual journals. Since journals relate to one another through citations and references (de Solla Price 1965), perhaps the best way to identify networked communities is through the cross-referencing of these already aggregated citations and references into algorithmically significant clusters (Leydesdorff 1987; Tijssen et al. 1987). This article describes a method of visualizing journal-to-journal connections to create ‘macro-epistemics’ (Knorr-Cetina 2007). New developments can be expected to form new journals and journal clusters (Van den Besselaar and Leydesdorff 1996).

The classification of journals into disciplines is complicated by the many venues where one finds results. Multidisciplinary journals such as Science and Nature play important roles in scientific communication, especially in calling attention to advances in knowledge. More recently, open-access journals (e.g., PLoS ONE) have emerged which deliberately ignore disciplinary boundaries and thus tend to disturb the classification of journals. Some scholars suggest that the journal form may diminish in use in favor of archives and repositories (Harnad 2001), although the majority of scholars view journals as increasingly important (e.g., Marbán 1999). As Lavoie et al. (2014) detail, “the transition from print to a digital, networked environment likely means that decision-making around the scholarly record will have to become more consciously coordinated”.

Journals are classified into disciplinary groups by indexing services; the classifications serve a number of purposes. First, classification serves to facilitate the process of search and retrieval. Secondly, bibliometric evaluations use journal classifications to normalize citation scores (Moed et al. 1995; Schubert and Braun 1986; Schubert et al. 1986). For pragmatic reasons, it has been considered “best practice” in evaluation studies to use the WoS Subject Categories (WCs)¹ for the operationalization of fields of science even though these categories do not represent homogeneous sets (Leydesdorff and Bornmann 2016). They are attributed to journals by manual indexing and have been elaborated incrementally for more than forty years by the providers of the database (Bensman and Leydesdorff 2009; Pudovkin and Garfield 2002, p. 1113). Journals can be attributed to more than one WC.

Beyond journal names and identity through sponsorship (e.g., by learned societies), articles can be classified in terms of co-citation, bibliographic coupling, or direct citation relations (Klavans and Boyack 2016, in press). Clustering the database at the level of papers, however, requires access to large computing capacity and to entire copies of Scopus (Boyack et al. 2011) or the WoS (Waltman and van Eck 2012). The problem of the validity of the delineations remains. As Schubert et al. (1989, at p. 7) have noted, “the field/subfield classification of papers is a neuralgic point of all kind of scientometric evaluations”.

Aware of the constraints of using WCs for evaluation purposes, Glänzel and Schubert (2003) developed a new journal classification system based on a pragmatic weighting of the results of algorithmic clustering of journals in terms of citation patterns against expert judgment. The Centre for Research and Development Monitoring ECOOM of the Catholic University of Leuven (Belgium) uses this classification system for evaluations. In the meantime, fast decomposition algorithms have been developed that can be used for classifications. Klavans and Boyack (in press, at p. 12, Table 3) list seven journal-based partitions of Scopus data currently in use.

Rafols and Leydesdorff (2009) compared (1) the WCs and (2) Glänzel and Schubert’s (2003) alternative classification with two algorithmically generated ones: (3) Newman and Girvan’s (2004) algorithm applied to the matrix of 7611 citing journals contained in the Journal Citation Reports (JCR) 2006; and (4) a random-walk based algorithm used by Rosvall and Bergstrom (2008) that had been applied to 6128 journals in the JCR 2004. The concordance between the four classifications was modest: in the 40–60% range (Rafols and Leydesdorff 2009, Table 3, at p. 1828). This conclusion agrees with Boyack’s estimate of 50% correct classifications for the WCs (Boyack, personal communication, 14 September 2008). However, most of the miscategorised journals appear to occur in areas within the close vicinity of categories indicated by the other classifications. In other words, the various decompositions are roughly consistent with each other, but imprecise. Despite the low correspondence, maps based on the different classifications can be rather similar (Leydesdorff and Rafols 2009; Klavans and Boyack 2009).

In summary, there are no unique or universally valid classifications of journals. Two runs of the same algorithmic decomposition may not provide the same results; most algorithms begin by drawing a random number using the computer clock. However, Leydesdorff et al. (2016b, at p. 907) noted that VOSviewer—visualization software developed by CWTS and available free for download at http://www.vosviewer.com—can generate quasi-deterministic classifications when the seed number of the randomizer is set equal to a constant (the default is zero). Using this option, the decomposition can pragmatically be combined with visualizations in a hierarchical classification by using the output of each decomposition recursively as input to the further decomposition at a next-lower level (Waltman et al. 2010). One begins at the top-level of the complete matrix and then extracts the clusters one by one; this process can be automated in a recursive loop if an option were added to VOSviewer for writing the output files to disk when running the program from the command line (Nees Jan van Eck, personal communication, 3 and 16 May 2016).

The most recent version 1.6.5 of VOSviewer (dated September 28, 2016), among other things, enables the user to run VOSviewer in a batch job from the command line. In this short communication, we report on generating such an automatic classification and visualization of the JCR-2015 data. The resulting classifications, visualizations, and routines are available at https://leydesdorff.github.io/journals/ and http://www.leydesdorff.net/jcr15. The website provides input files for journal maps for the more than 11,000 journals contained in the JCR-2015, at the various levels of clustering.

Although developed for JCR-data, the routines are formulated so that any 1-mode matrix can be decomposed similarly in terms of mappings using VOSviewer. Note that one can also export the clusters in the Pajek format so that the files can be used for other visualizations such as in Gephi.

The matrix of aggregated journal–journal citation relations represents a complex system of scientific communication that is both hierarchically layered and functionally differentiated in terms of scientific specialties and fields. Such a system cannot be decomposed unambiguously (Simon 1973). The clusters can be related in other (e.g., methodological versus theoretical) dimensions; densities of communication in subsets can vary significantly. Referencing behavior norms may differ across fields. For example, an article in a biomedical specialty may contain more than forty references, while in other fields, such as mathematics, fewer than ten references is more common (Garfield 1979; Moed 2010). However, what is measured as “differences in citation behavior among fields” can also be an artifact of the different degrees of coverage of the field-specific literature by the database (Marx and Bornmann 2015). Epistemically, references may function at research fronts to position the citing papers or acknowledge intellectual debt and/or credit to previous (that is, cited) publications (Leydesdorff et al. 2016a). Bodies of specialist literature may interact in next-order—i.e., more generalist—layers carried by quality journals such as Science and Nature.

This complex interweaving of different dynamics is further complicated because all relevant distributions are heavily skewed (Seglen 1992). Weak ties in one context can be strong ties from another perspective (Granovetter 1973). As we have seen above, hierarchical decomposition follows a path downward so that the results are path-dependent and may lead to different sub-optima. There is no objective yardstick to inform us how much better one representation is when compared with another (cf. Klavans and Boyack 2016, in press).

In addition to the statistical quality of the distinctions, the groupings have to be labeled; this adds a subjective dimension of flexible interpretations with different meanings, since the labels are not provided by the decomposition itself. The labels are added by an analyst who, as a user of the system, may wish to mix pragmatic with intellectual considerations. Ex ante, one representation is as legitimate as another and no methodological prescription can be formulated.

Within this context of uncertainty and complexity, the proposed routine provides a means for testing one’s assumptions without claiming authority; but with the advantage of reproducibility and the possibility of rich visualizations. The algorithm is semantically neutral: the routine will work on any 1-mode matrix and provide a purely algorithmic decomposition of the system into lower-level units in a series of layers. The advantages of using this decomposition and the quality of the visualizations will have to show their usefulness in bibliometric practices. The results may raise further questions and thus help to shape research ideas and agendas.