Granularity of algorithmically constructed publication-level classifications of research publications: Identification of specialties

In a recent article we proposed a methodology for identification of research topics in an algorithmically constructed publication-level classification of research publications (ACPLC; Sjögårde & Ahlgren, 2018). We used a large dataset of more than 30 million publications in Web of Science to create an ACPLC, at the granularity level of topics. We consider topics as problem areas addressed by researchers, representing “an underlying semantic theme” (Yan et al., 2012), and we see topics as the lowest level of aggregation to be considered for classification of subject areas (Besselaar & Heimeriks, 2006). However, more levels of different granularity are needed for an ACPLC to be used to answer a broader range of questions. In the present study, we use a similar methodology to create a classification whose granularity corresponds to research specialties. In the remainder of this paper, we use the term “specialty” instead of “research specialty.” In short, a specialty is a “network of researchers who tend to study the same research topics” (Morris & Van der Veer Martens, 2008). However, the specialty notion is further discussed below. In this paper we identify the publications belonging to specialties by grouping the topics obtained in the previous study.

The identification of specialties is part of a broader aim to develop a standard approach to create a large and global hierarchical ACPLC of research publications in terms of geographical uptake, coverage of subject areas, and citation databases, such as Web of Science or Scopus. An ACPLC can be used for a great variety of analytical purposes and is especially useful for recurrent analytical activities.

A classification system that groups publications into classes whose sizes correspond to specialties can be used to study the publication output of different actors within a specialty; the collaboration between actors, dynamics, emergence and decline of specialties; and the relation between specialties. Moreover, a hierarchical classification, including both classes corresponding to topics and classes corresponding to specialties, makes it possible to identify topics within a specialty and, for example, a shifting focus of a specialty. We therefore suggest that the level of specialties, together with the level of topics, should be included in a standard ACPLC, and that such an ACPLC should be hierarchical.

The purpose of this paper is to find a theoretically grounded, practically applicable, and useful granularity level of an ACPLC with respect to specialties. To determine the granularity of specialties, a baseline classification is constructed. A set of journals is identified and used to create a baseline classification. ACPLCs with different granularities, constructed by the use of different values of a resolution parameter, are then compared to the baseline classification. The classification that best fits the baseline classification is proposed to be used for bibliometric analyses of specialties. In contrast to earlier work, our aim is to create a classification of publications that can be used to identify all specialties represented in Web of Science from 1980 onwards.

The remainder of this paper is structured as follows. In the next section, a short summary of our previous article on topic identification is given. The framework of the study is outlined in Section 3 and the specialty notion is discussed in Section 4. Data and methods are presented in Section 5, and Section 6 gives the results. Conclusions are given in Section 7.

To give the reader some background to the present study, in this section we summarize the earlier study on topic identification (Sjögårde & Ahlgren, 2018). In that study, we discussed how the resolution parameter given to the software Modularity Optimizer can be calibrated to obtain publication classes corresponding to the size of topics.

A set of about 31 million articles and reviews from Bibmet, KTH Royal Institute of Technology’s bibliometric database, which contains Web of Science data, was used for the study. The study involved a methodology consisting of four steps. In the first step, we constructed a baseline classification (BCP_t) corresponding to topics, where BCP_t contains synthesis articles, operationalized as articles with at least 100 references. Each such article constitutes a class, and its list of cited references points to the reference articles of the class (i.e., to the members of the class). The underlying assumption of this approach is that synthesis publications in general address a topic.

In the second step of the methodology, several ACPLCs of different granularity with respect to the topic level were created by setting the resolution parameter of Modularity Optimizer to different values. Normalized direct citation values between the articles in the dataset were used, as proposed by Waltman and van Eck (2012). For the third step, classifications derived from the ACPLCs were obtained, where each derived classification constitutes a classification of the union of the classes of the baseline classification, BCP_t. Thus, the latter classification and a given derived one have exactly the same underlying reference articles. In the fourth and final step of the methodology, the similarity between BCP_t and each of the derived classifications from the third step was quantified. For this purpose, the Adjusted Rand Index (ARI; Hubert & Arabie, 1985) was used. We denoted the ACPLC such that its corresponding derived classification exhibited the largest ARI similarity with BCP_t by ACPLC_t.

With respect to the results of the study, the class size variation regarding ACPLC_t turned out to be moderate, and only a small proportion of the articles belong to very small classes. Moreover, the outcomes of two case studies showed that the topics of the cases were closely associated with different classes of ACPLC_t, and that these classes tend to treat only one topic. We concluded that the proposed methodology is suitable to determine the topic granularity level of an ACPLC and that the ACPLC identified by this methodology is useful for bibliometric analyses.

In the present study, we use a similar methodology to identify specialties. The 230,559 classes obtained in the previous study, of which 136,939 have a size of at least 50 articles, are clustered into specialties. A baseline classification is constructed that corresponds to specialties, and a set of journals is used to create the baseline classification.

We need to point out that there is a substantial overlap between our earlier paper (Sjögårde & Ahlgren, 2018) and the present one. The reason for this is that the four-step methodology used in the earlier study, and briefly described above, is also used in the study underlying the present paper.

As in the previous study, we use a network-based approach to obtain a classification of research publications (Fortunato, 2010). We use the Modularity Optimizer¹ software, created by Waltman and van Eck (2013), and the methodology put forward in Waltman and van Eck (2012). This framework has also been used by others (Klavans & Boyack, 2017a, b). The alternative modularity function is used (Traag et al., 2011), together with the SLM algorithm for modularity optimization. We acknowledge that a new algorithm for modularity optimization has been proposed (Traag et al., 2019). However, to be consistent with the previous study, we use the SLM algorithm in this study. We choose direct citation to express publication-publication relations, rather than bibliographic coupling (Kessler, 1965), cocitations (e.g., Marshakova-Shaikevich, 1973; Small, 1974), textual similarity (e.g., Ahlgren & Colliander, 2009; Boyack et al., 2011), or combined approaches (e.g., Colliander, 2015; Glänzel & Thijs, 2017). Direct citation is more efficient as it gives rise to fewer relations than the mentioned approaches, and there is empirical support that direct citations perform well in comparison with bibliographic coupling and cocitations when it comes to larger data sets (Boyack, 2017).

In Sjögårde and Ahlgren (2018), a network model with two levels of hierarchy, topics and specialties, was presented. This model comprises a logical classification: Each publication is classified into exactly one class at each level of hierarchy.² Moreover, all publications in a class, at a level below the top level, are classified into exactly one and the same parent class. It follows that each topic in the model belongs to exactly one specialty. In this study, in which we continue to use logical classifications, we obtain such a relation by clustering topics into specialties, rather than using the alternative approach to cluster publications directly into specialties. Logical classifications have some shortcomings: Topics can be addressed by several specialties (Yan et al., 2012) or, at a higher level of aggregation, disciplines (Wen et al., 2017), phenomena not expressed by logical classifications. However, the relation between a topic and other specialties than the parent specialty, as well as relations between topics, can still be expressed and analyzed by use of the relational strengths associated with the edges in the model.

For further discussion on the general classification framework and for an explication of a model that expresses the relations between classes at different hierarchical levels in the model, we refer the reader to Sjögårde and Ahlgren (2018).

Specialties have been studied since the 1960s in the field of sociology. In this literature, specialties are considered as smaller intellectual units within research disciplines (Chubin, 1976). The researchers within the same specialty communicate with each other. They possess similar competences and can engage in the same, or similar, research problems (Hagstrom, 1970). The notion of specialties is closely related to the notion of invisible colleges (Crane, 1972; Price, 1965). However, as pointed out by Morris and van der Veer Martens (2008), invisible colleges “presuppose that the researchers are in frequent informal contact with one another,” which is not the case for specialties.

We use the definition of a specialty that has been given by Morris and van der Veer Martens (2008). They define a specialty as “a self-organized network of researchers who tend to study the same research topics, attend the same conferences, read and cite each other’s research papers and publish in the same journals.” Further, and in concurrence with others, we consider specialties to be the largest homogeneous units of science “in that each specialty has its own set of problems, a core of researchers, shared knowledge, a vocabulary, and literature” (Scharnhorst et al., 2012) and that they “play an important role in the creation and validation of new knowledge” (Colliander, 2014).

As early as 1974, Small and Griffith argued that publications can be clustered and that the obtained clusters may represent specialties (Small & Griffith, 1974). The single-linkage method was used by Small and Griffith to cluster 1,832 publications, which today would be considered a very small number of publications. They used their results to identify specialties. Since the 1970s, the technological advancements and the emergence of the Internet have changed the preconditions for research communication. There has also been a growth in research activity and production of research publications.

More lately, specialties have been identified and analyzed by the use of different clustering techniques (Lucio-Arias & Leydesdorff, 2009; Morris & van der Veer Martens, 2008; Scharnhorst et al., 2012). Different points of departure and different operationalizations of the specialty notion have captured different aspects of specialties. For example, clustering of publications based on citation relations and clustering of researchers based on coauthorship may result in different pictures of a specialty. The former approach identifies a set of publications and the latter a group of researchers belonging to a specialty. We attempt to capture the publications belonging to each specialty, rather than the researchers belonging to the specialty. A researcher can be part of several specialties, a property that cannot be expressed by the coauthorship approach. For this reason, we consider this approach less suitable for the identification of publications belonging to a specialty. We believe that it is preferable to base classifications constructed for the purpose of bibliometric analyses of specialties on the network of publications, rather than on the network of researchers. Our approach makes it possible to identify the researchers within a specialty without forcing every researcher into exactly one specialty. It also makes it possible to analyze the contribution of one researcher to multiple specialties.

Kuhn (1996) estimates the number of core researchers in a specialty to be around 100. Based on Lotka’s law (1926), Morris (2005) estimates the total number of researchers within a specialty to be around 1,000, and the number of publications produced by a specialty to be between 100 and 5,000. Boyack et al. (2014) regard specialties to be “ranging from roughly a hundred to a thousand articles per year.” We acknowledge that the size of specialties in terms of publications may vary over time. Because the output of research publications has been growing the last decades, it is likely that the total size of specialties, in terms of number of publications, has been growing. Also, the yearly publication production of active specialties is likely to be on average larger today than 10 or 20 years ago. The size of specialties is an empirical question that we intend to shed light on in the present study.

As in Sjögårde and Ahlgren (2018), KTH Royal Institute of Technology’s bibliometric database Bibmet was used for the study. Bibmet contains Web of Science publications from the publication year 1980 onwards. In the present study, we use the same set of publications as in the earlier study. We denote this set, in agreement with the earlier study, by P. P consists of 30,669,365 publications of the two document types: “Article” and “Review.” In the remainder of this paper, we use the term “article” to refer to both articles and reviews.

We attempt to find a granularity of an ACPLC, where the ACPLC is based on the articles in P, that corresponds to specialties. In order to identify the granularity of specialties, a baseline classification of publications (BCP) is created. The BCP is a set of journals, considered as classes, and each member of a class in BCP is a publication appearing in the class (i.e., appearing in the journal).

The BCP is compared to several ACPLCs with different granularities, where each such ACPLC is obtained by clustering the classes of ACPLC_t (see Section 2), which is thereby utilized in the present study. An appropriate granularity is detected and an ACPLC is chosen, the classes of which correspond to specialties. The methodology, which has four steps and a high degree of similarity with the methodology proposed in Sjögårde and Ahlgren (2018), is described in detail in steps I to IV below and schematically illustrated in Figure 1.

We construct a baseline classification to correspond to specialties, which we denote by BCP_s. For the creation of BCP_s, a subset of journals covered by Web of Science is used. Each journal constitutes a class, and the publications appearing in the journal are the members of the class.

The reason for using journals to obtain BCP_s is that researchers within a specialty publish in and read the same journals. The new possibilities to search, retrieve and read research articles have changed the role of journals, but nevertheless many journals are still focused on specific areas of expertise and the researchers within those areas. Such journals aim to publish articles that are relevant to its audience. For example, we consider bibliometrics as a specialty within the discipline of library and information science, and the scope of the Journal of Informetrics as roughly targeting the specialty of bibliometrics. In resemblance with Bradford’s law (1948), researchers within a specialty need to go to several journals to find all the relevant articles within their specialty. The boundaries of a specialty are vague and fading rather than sharp. If we consider a journal, the scope of which roughly covers a specialty, a core set of the articles in such journal is likely to be of high relevance to the core audience of the journal. The researchers that belong to this core audience can be considered as the backbone of the specialty. The rest of the articles in the journal have a fading relevance to this specialty. Some of these articles will be of higher relevance to other specialties.

When creating BCP_s, we attempt to delimit the set of journals to such journals that, regarding their size and scope, can be considered as proxies for specialties. As BCP_s is to be used as a baseline to estimate the granularity of an ACPLC regarding specialties, the following three requirements should be addressed:

• A. 

  To be able to compare the classifications, the union of the classes in BCP_s must be a subset of the union of the classes (i.e., the topics) in ACPLC_t.

• B. 

  Ideally, each class (journal) in BCP_s should address exactly one specialty.

• C. 

  Ideally, each pair of distinct classes (journals) should address different specialties.

Now, to satisfy point A, we kept, for a given journal, only articles (i.e., publications that are of the document types “Article” or “Review”) that are present in ACPLC_t (i.e., having a classification at the topic level).

To deal with point B, we first delimited the publication period to five years, namely 2008–2012. By this operation, which resulted in 6,140,762 publications in 13,070 journals, the risk of including journals that, for instance, have shifted subject focus over time is lowered. In addition to dealing with point B, the choice of publications from publication years that have both incoming and outgoing citations can be assumed to have a stabilizing effect when these articles are being clustered, compared to more recent publications.

We then removed all journals belonging to the Web of Science subject category “Multidisciplinary Sciences,” because a journal in this category is clearly not focused on a single specialty. After this, 13,023 journals remained. Next, we considered the distribution of journals by size. Figure 2 shows the distribution limited to journals with less than or equal to 2,000 articles. A typical journal, with respect to size and modal interval as a measure of central tendency, published 90–100 articles from 2008 to 2012. By including journals between the 10th and 75th percentiles of the journal size distribution displayed in Figure 2, journals with 47–478 articles were included. With this journal size limitation, the risk to include journals addressing multiple specialties (or journals with a narrower scope than a specialty) is reduced. The limitation reduced the number of journals to 8,485.

Finally, in order to further reduce the risk of including journals addressing multiple specialties, we took journal self-citations into account. The idea is that a one-specialty journal can be assumed to cite itself to a larger extent compared to a journal that covers two or more specialties, other things held constant. In the light of this, we required, for a journal to be included in BCP_s, that the self-citation ratio (in %) should be at least 10.³ The journal set was reduced to 1,540 journals by this procedure. Some test runs with different values of the threshold were conducted. These runs showed that lower values of the threshold reduced the maximum ARI value (cf. step IV below), which indicates that lowering the threshold value results in broader, less focused journals. The threshold was set to include as many journals as possible and to keep the ARI value reasonably high.

Some of the measures taken to satisfy point B are also relevant for satisfying point C (which states that each pair of distinct classes should address different specialties), such as the limitation to the publication years 2008–2012. With the aim to further raise the possibilities of satisfying point C, we applied bibliographic coupling between journals. If two journals had an overlap of 8% or more regarding their active cited references, they were considered as specialty overlapping.⁴ This threshold was chosen after browsing the list of journal pairs sorted in descending order based on number of shared cited references. Based on journal titles, it is obvious that some journal pairs have an overlapping subject focus: for example, the two journals Higher Education and Studies in Higher Education (19% citation overlap). A threshold for the cited references overlap was chosen to include such apparent cases. In addition, test runs were conducted with different threshold values. Higher values resulted in lower maximum ARI values. For this reason, we tried to keep the threshold value as low as possible (without considering journals with nonoverlapping subject focus as specialty overlapping).

We grouped journals so that all journals that were directly or indirectly connected, by a cited reference overlap of 8% or more, were assigned the same group. For example, if journal j₁ has a cited reference overlap of ≥ 8% with journal j₂, and j₂ has a cited reference overlap of ≥ 8% with j₃, then j₁, j₂, and j₃ are assigned to the same group. Note that j₁ and j₃ are assigned to the same group even if they do not have an active reference article overlap of ≥ 8%. Each obtained group of journals was considered as addressing the same specialty. One of the journals was then randomly selected from each group. After the execution of this procedure, 967 journals remained. This number is the number of journals (classes) in BCP_s. We denote the union of the classes in BCP_s as P′.

In order to obtain ACPLCs of different granularity, the first step was to measure the relatedness between the classes (topics) of ACPLC_t. We measured the relatedness as the average normalized direct citation value between the articles belonging to the two classes: If class C contains m articles and class C′ n, the sum of the m × n normalized direct citation values between articles in C and articles in C′ was divided by m × n. In the second step, the generated class relatedness values were iteratively given as input to Modularity Optimizer to cluster the classes of ACPLC_t, where the resolution parameter was set to different values in the iterations.⁵ By this, ACPLCs were created for comparison of similarity with BCP_s. We denote the ACPLCs by ACPLC_1, …, ACPLC_k, where k is the number of created ACPLCs.

For each ACPLC_i (1 ≤ i ≤ k), a classification was derived from ACPLC_i in the following way:

• (a) 

  Each class C in ACPLC_i such that C did not contain any articles in P′ was removed from ACPLC_i. Let ACPLC_i1 be the subset of ACPLC_i that resulted from the removal.

• (b) 

  For each class C in ACPLC_i1, all articles in C that did not belong to P′ were removed from C. Let ACPLC_iP′ be the set that resulted from these removal operations.

Clearly, the set ACPLC_iP′ constitutes a classification of P′ (i.e., of the union of the classes of the baseline classification BCP_s). Thus, ACPLC_iP′ and BCP_s have exactly the same underlying articles. We denote the k derived classifications as ACPLC_1P′, …, ACPLC_kP′. These classifications then correspond to the classifications ACPLC_1, …, ACPLC_k.

We attempt to optimize the granularity of an ACPLC_iP′ so that it exhibits as high a similarity as possible with BCP_s. Figure 3 illustrates the relation between two classifications as an alluvial diagram. Example A shows two classifications A₁ and A₂ with a high similarity. Example B shows two classifications where one of the classifications is more coarsely grained (B₁) than the other classification (B₂). The similarity between A₁ and A₂ is higher than the similarity between B₁ and B₂. If we consider B₁ as a baseline classification, then the granularity of B₂ would be too finely grained.

As in our topic identification study, we used the ARI (Hubert & Arabie, 1985) to quantify the similarity between BCP_s and an ACPLC_iP′. The ARI ranges from 0 to 1. It is advantageous over the original Rand Index proposed by Rand (1971), because it adjusts for chance. The ARI compares two classifications by considering pairs of items in one of the classifications and whether or not each pair is grouped into the same class in the other classification. Note that an ARI value of 1 between BCP_t and an ACPLC_iP′ corresponds to a situation in which these two classifications are identical. For further information on ARI, we refer the reader to Sjögårde and Ahlgren (2018).

To find the ACPLC_iP′ with the highest ARI similarity with BCP_s, we tested the similarity after each run of Modularity Optimizer. A first run was made with a resolution parameter value of 5E-7. This value was chosen based on previous experience and some testing. We then increased the parameter value by 5E-7. This increase resulted in a higher ARI similarity, and we therefore increased the resolution further by 5E-7 for the third run, from 1E-6 to 1.5E-6. We continued by increasing the resolution by 5E-7 in total four more times, and thus seven runs were done. The fifth run, with a resolution parameter value of 2.5E-6, gave rise to the highest ARI similarity (see Table 2 and Figure 4, Section 6).

In total BCP_s consists of 967 baseline classes. A given ACLPC_iP′ consists of 202,647 articles, which is about 3.3% of the articles from the years 2008–2012 in the corresponding ACPLC_i. The ACPLC_i such that ACLPC_iP′ exhibits the largest ARI similarity with BCP_s is proposed to be used for the analyses of specialties. We denote this ACPLC_i by ACPLC_s.

In this section, we first deal with the selection and properties of ACPLC_s. Then, as in the earlier study on topic identification (Sjögårde & Ahlgren, 2018), we consider two cases. We examine the specialties of articles belonging to (1) the Web of Science subject category “Information science & Library Science,” and (2) the Web of Science subject category “Medical Informatics.”

Figure 4 shows a scatterplot of the relation between the resolution value (horizontal axis) used to obtain ACPLC_is and the ARI value (vertical axis), obtained by comparing the ACPLC_iP′s with BCP_s. ACPLC_5P′ has the highest ARI value. ACPLC_5P′ corresponds to ACPLC_5, which we consider to be the most proper ACPLC_i with respect to granularity of specialties. In the remainder of this paper, we denote ACPLC_5 as ACPLC_s. However, we acknowledge that ACPLC_4P′ and ACPLC_6P′ have ARI values that are only slightly lower/higher than the value of ACPLC_5P′. Thus, ACPLC_4P′ and ACPLC_6P′ perform almost as well as ACPLC_5P′.

To get a picture of how well ACPLC_s matches BCP_s, we calculated the distribution of articles in an average class in BCP_s into classes (journals) in ACPLC_s. This was done by first calculating the average number of classes in ACPLC_s into which the articles in a class in BCP_s are distributed, an average that is equal to 50 (after rounding to the nearest integer). We then selected all 12 classes in BCP_s that were distributed into exactly 50 classes. Let the set of these classes be P_sc. The average number of articles in a P_sc class is 160.3. For each of the P_sc classes, we calculated the number of its articles in each of the 50 ACPLC_s classes and sorted the resulting table in descending order. The ACPLC_s class with the highest number of articles (i.e., the class corresponding to the first row in the table) was assigned rank 1, the second largest class (i.e., the class corresponding to the second row in the table) was assigned rank 2, etc. In this way, 12 ranked tables were obtained. Finally, averages of the number of articles by rank number, 1, …, 50, were calculated across all the 12 tables. Figure 5 shows the resulting average distribution of articles in P_sc (to the left) into the 50 ACPLC_s classes (to the right). Ranks and average number of articles across the P_sc classes are shown for ACPLC_s.

Given that we consider the classes in ACPLC_s as specialties, the distribution of journal articles in a typical BCP_s class follows a skewed distribution of specialties. About 41% of the articles in an average BCP_s class are distributed into the two most frequent specialties, and 34 specialties (classes 17 to 50) are represented by a single article (after rounding to nearest integer). Hence, a high share of the articles of the average BCP_s class is concentrated to a few of the ACPLC_s classes. We therefore consider the match between ACPLC_s and BCP_s as good.

ACPLC_s consists of 61,805 classes, ranging from 1 to 46,078 articles. Most of the classes are small in size; however, these classes contain a small share of the total number of articles in ACPLC_s. For instance, classes with fewer than 500 articles contain about 1.2% of the articles in ACPLC_s. Figure 6 shows a histogram of the distribution of classes by class size (in terms of number of articles). In order to include classes of a substantial size in the figure, classes with fewer than 500 articles have been excluded from the figure.

Most specialties of substantial size (minimum of 500 articles) have 5 (10th percentile) to 62 (90th percentile) subordinated topics of substantial size (a minimum of 50 articles), with a mode of 6, a median of 19 and a mean of about 28 (Figure 7 and Table 1).

In Figure 8, class sizes are plotted by rank order for ACPLC_s (= ACPLC_5), as well as for ACPLC_4 and ACPLC_6. A log-10 scale is used on both the vertical axis (showing class sizes by number of articles) and the horizontal axis (showing ranks). In this figure, all classes are shown, including small size classes. About 4,200 classes contain at least 1,000 articles, about 1,000 classes contain at least 10,000 articles and about 30 classes contain at least 30,000 articles. In agreement with our study on topics, the size of classes is dropping rather slowly, regardless of classification. The increasing granularity—from ACPLC_4 via ACPLC_s to ACPLC_6—is reflected by, for example, corresponding, decreasing intercepts.

Figure 9 expresses the number of articles in P (vertical axis) that is associated with different class sizes (horizontal axis). For a randomly selected article a, it is most probable that the size of the specialty class in ACPLC_s to which a belongs is 6,000–7,000 articles (cf. the highest bar of the histogram in Figure 9). Eighty percent of the articles belong to classes consisting of 2,899 (10th percentile) to 22,819 (90th percentile) articles (Table 2). The median value of ACPLC_s is 10,499 and the mean 12,016. This distribution is not as skewed as the corresponding topic distribution (Sjögårde & Ahlgren, 2018, Figure 8).

The number of articles contributing to a specialty in 2015 (the most recent complete year at the time for data extraction) is between 148 and 1,597, given that we only take the mid-80% of the distribution into account (Table 3 and Figure 10). The median class size is 593. The mean number of articles per specialty class is growing approximately linearly across the 10-year period (Table 3). This can be expected, considering the linear growth of research publications in Web of Science.

As mentioned in the introduction, Morris (2005) estimates the size of specialties to be between 100 and 5,000 articles (but not mentioning any time period), and Boyack et al. (2014) estimate the yearly article output of a specialty to be somewhere between 100 and 1,000 articles. The results of the present study cannot be easily compared to these figures. The estimations of Morris and Boyack et al. are rough. Morris does not mention any time period. Further, the work by Morris is rather old and the size of specialties may have increased in terms of publication output. Table 3 shows that the number of articles in Web of Science has been growing by more than 50% between 2006 and 2015. In 2015, the size of specialties ranges from about 150 articles (10th percentile) to 1,600 (90th percentile) articles. Thus, the size of specialties in 2015 is about 50% larger than the size estimated by Boyack et al. We regard this difference as rather small, taking into account that Boyack et al. define the next larger level (disciplines) to range from tens to hundreds of thousands of articles per year, several orders of magnitude larger than our estimation of the size of specialties.

In agreement with Morris and Boyack et al., we find it reasonable not to consider publication classes under some threshold to be regarded as specialties. One solution to the problem of small class sizes is to reassign such classes (classes below a threshold) based on their relations with larger classes (classes above or equal to the same threshold) as proposed by Waltman and van Eck (2012). However, how to set the threshold is a question that we do not address in this paper.

To explore how articles within the discipline of library and information science (LIS) are distributed into classes in ACPLC_s, we retrieved all articles in P that belong to a journal classified into the Web of Science subject category “Information Science & Library Science” and published in the period 2011–2015. In total, 16,278 articles were retrieved. Let P_lis be this set of articles.

For each class in ACPLC_s, labels were automatically created based on author keywords. Chi-square was used to quantify the relevance of author keywords in each class, and for each class, three author keywords with highest rank were concatenated to a label (for more detail see Sjögårde & Ahlgren, 2018). To distinguish the scope of each specialty, we used these labels and the labels of the topics in each class. Recall that ACPLC_s is obtained by clustering the topics of ACPLC_t, the best performing ACPLC with respect to topic identification (Sjögårde & Ahlgren, 2018).

Table 4 shows the total number of articles in the 10 most frequent specialties and the number, and the share, of articles in a specialty that belong to P_lis. The top 10 specialties cover about 48% of the articles in P_lis. Some of the top 10 specialties are highly concentrated within the analyzed Web of Science subject category (e.g., “INFORMATION LITERACY//PUBLIC LIBRARIES//ACADEMIC LIBRARIES,” 79%), whereas other specialties have a low share of its total number of articles in this category (e.g., “INNOVATION//PATENTS//OPEN INNOVATION,” 7%).

The highest ranked specialty, “BIBLIOMETRICS//CITATION ANALYSIS//IMPACT FACTOR,” focuses on bibliometric indicators, mapping and evaluation of research, and the analysis of scholarly communication. We acknowledge that a majority of the largest topics in this specialty are the same topics that were observed in the case study of Journal of Informetrics in the previous topics study (Sjögårde & Ahlgren, 2018, and  Appendix 1 in this paper). The second-ranked specialty, “INFORMATION LITERACY//PUBLIC LIBRARIES//ACADEMIC LIBRARIES,” focuses on library science. This category includes topics such as information literacy, knowledge organization, information practices and reference services. The specialty “INTERLENDING//DOCUMENT DELIVERY//ACADEMIC LIBRARIES” includes topics specifically related to academic libraries, such as electronic media, open access, interlending, library circulation systems and data repositories. The scopes of specialties 4, 6, 7, 8 and 10 are captured rather well by their labels, and these specialties are all clearly related to LIS. These five specialties include information retrieval, knowledge management, library and information aspects of health service and occupation as well as of innovation and patents. The specialty “ENTERPRISE RESOURCE PLANNING//ENTERPRISE RESOURCE PLANNING ERP//END USER COMPUTING” includes some topics related to LIS (e.g., IT business value, IT outsourcing, Information system planning, and Information infrastructure). The LIS relevance of “UNIVERSAL SERVICE//TELECOMMUNICATIONS//ACCESS PRICING” (rank 9) is within topics such as Internet access and Digital divide.

Appendix 1 lists the 10 topics with most publications in P_lis for the top 10 ranked specialties with regard to P_lis.

In analogy with the case of LIS, we retrieved all articles in P that belong to a Web of Science subject category, in this case “Medical Informatics,” and published in the period 2011–2015, to explore how articles within this discipline are distributed into classes in ACPLC_s. In total, 12,516 articles were retrieved. Let P_mi be this set of articles.

Table 5 shows the top 10 specialties in P_mi, ranked by frequency. Only one specialty is highly concentrated into the “Medical Informatics” category, namely “ELECTRONIC HEALTH RECORDS//ELECTRONIC MEDICAL RECORD//MEDICAL INFORMATICS” (which is also present in the LIS case). For the rest of the top 10 specialties, 14% or less of the articles in the specialty belong to P_mi. This might suggest that MI is more interdisciplinary than LIS. It can also be the case that MI articles are published in broader journals, which are not classified into the “Medical Informatics” Web of Science subject category.

The largest specialty in the “Medical Informatics” category focuses on clinical decision support systems, clinical research informatics and electronic health records. The second-ranked specialty within the category, “HEALTH LITERACY//INTERNET//MHEALTH,” addresses topics within mobile health such as personal health records, online health information, and online support groups. The specialty “HEALTH TECHNOLOGY ASSESSMENT//EQ 5D//PRIORITY SETTING” focuses on health technology assessment and cost effectiveness.

The remaining seven top 10 ranked specialties have the following foci: (4) clinical trial designs; (5) mathematical and statistical models and methods within the medical sciences; (6) prediction and risk models; (7) evidence-based medicine, medical epistemology, meta-analysis methods, and literature searching; (8) Patient safety (includes incident and error reporting); (9) telehealth (can be seen as a predecessor to mobile health); and (10) gene ontologies.

Appendix 2 lists the 10 topics with most publications in P_mi for the top 10 ranked specialties with regard to P_mi.

In this study we have discussed how the resolution parameter given to the Modularity Optimizer software can be calibrated to cluster topics, obtained in a previous study on topic identification (Sjögårde & Ahlgren, 2018), so that the obtained publication classes correspond to the size of specialties. A set of journals has been used as baseline for the calibration. Journals were selected based on their size and self-citation rate. The underlying assumption of our approach is that journals of a particular size and focus have a scope that corresponds to specialties. By measuring the similarity between (1) the baseline classification and (2) multiple classifications obtained by using different values of the resolution parameter, we have identified a classification, which we denote as ACPLC_s, whose granularity corresponds to specialties.

Some criteria for the evaluation of ACPLC_t, the best performing ACPLC with respect to topic identification, are the same for the evaluation of ACPLC_s. The differences in class sizes should not be too large and “the number of very small clusters should be minimized as much as possible” (Šubelj et al., 2016). In ACPLC_s, 80% of the articles belong to classes consisting of 2,899–22,819 articles. Further, 80% of the articles belong to classes with a yearly publication rate of 98–869 articles in publication year 2006, increasing to 148–1,597 in the publication year 2015. Only 1.2% of the articles in ACPLC_s belong to classes with a total number of articles less than 500. As in the previous study, the distribution follows a typical scientometric distribution, and we therefore consider the results, regarding class sizes, as satisfying.

In the present study, we have not implemented a reclassification of small classes. However, in accordance with the previous study, we consider reclassification of small classes to be desirable for practical reasons. Moreover, we think that content labeling of classes is a topic for future work.

Another criterion stated by Šubelj et al. (2016) is that classes should make intuitive sense. In addition, we stress that the focus of a specialty should be possible to identify and that two specialties should have subject foci that can be distinguished. Two case studies, in which we have identified specialties within the disciplines of LIS and MI, have been performed to evaluate these criteria. We could identify the subject foci of the specialties in these case studies, and the subject foci of the specialties have been relatively easy to distinguish. Thus, the two criteria are (approximately) satisfied in our case. Further, several of the specialties identified in the LIS case have been identified by others (Bauer et al., 2016; Blessinger & Frasier, n.d.; Figuerola et al., 2017; Janssens et al., 2006) and the same holds for several of the specialties identified in the MI case (Kim & Delen, 2018; Schuemie et al., 2009; Wang et al., 2017). However, more case studies are needed to verify the soundness of the methodology used.

The aforementioned feature of the classification approach used in this study, logical classification, which assigns each topic to exactly one speciality, has some limitations. It is clear that topics can be addressed by several specialties (or at higher level disciplines). For instance,  Appendix 1 and 2 show that the topic with the label “NATURAL LANGUAGE PROCESSING//MEDICAL LANGUAGE PROCESSING//CLINICAL TEXT” is addressed by both the LIS and MI disciplines. This topic is forced into exactly one specialty, “ELECTRONIC HEALTH RECORDS//ELECTRONIC MEDICAL RECORD//MEDICAL INFORMATICS.” Thus, relations between this topic and, for example, specialties within the LIS discipline are not expressed by ACPLC_s. However, relations between a specialty and topics within other specialties can still be analyzed using, for instance, citation relations. Nevertheless, a logical classification to some extent oversimplifies the complex structure of topic representation in research publications.

We acknowledge that direct citations perform less well than bibliographic coupling in a recent study (Waltman et al., 2019). However, a relatively low number of articles was used in the study, about 700,000 in comparison with the over 31 million articles used in this study. Moreover, a relatively short publication window (2007–2016), in comparison with the present study (1980–2016), was used. Interestingly, the study shows that an extended direct citation approach, in which direct citation relations within an extended set of publications are taken into account, performs better than an ordinary direct citation approach. Which publication-publication similarity measure to be used for the creation of an ACPLC still needs to be further investigated, however.

We recognize that there is only a small difference in performance, regarding the ARI values, between ACPLC_4P′, ACPLC_5P′ and ACPLC_6P′.⁷ Therefore, we can only determine the granularity of specialties roughly. This is, however, not surprising, given the complex, overlapping nature of research subject areas. Nevertheless, this study sets a benchmark of the size of specialties and outlines a methodology for the calibration of ACPLCs.

The combined outcome of our previous study on the classification of topics, and the present study on the classification of specialties, is a two-level hierarchical classification. We believe that such a classification comprises a valuable part of a research information system and propose that such a classification can be used for bibliometric analyses of topics and specialties.

Peter Sjögårde: Conceptualization; methodology; software; formal analysis; writing—original draft; writing—review & editing; visualization. Per Ahlgren: Conceptualization; methodology; formal analysis; writing—original draft; writing—review & editing.

The authors have no competing interests.

No funding has been received.

The data analyzed in this manuscript is subject to copyright (by Clarivate Analytics^®, Philadelphia, Pennsylvania, USA) and cannot be made available.

We would like to thank two anonymous reviewers for their relevant and constructive comments on an earlier version of this paper.