Research Topic Specialization of Universities in Information Science and Library Science and Its Impact on Inter-University Collaboration

Abstract

Universities significantly empower the development of science and technology, and inter-university research collaborations have been one of the major approaches. Considering each university has its expertise regarding research topics in a given discipline, the present paper examines the specialization of university research and its impact on inter-university collaborations. Based on a keyword network constructed via research articles in Information Science and Library Science, 10 research topics are identified. Accordingly, the research topic diversity of a university and the research topic similarity between two universities are quantified. The universities with diverse research topics are found to be more collaborative. A further collaboration network analysis based on the Quadradic Assignment Procedure reveals the important role of research topic similarity on the closeness and impact of collaborations. The different research topic specializations largely prevent close collaborations between two universities, but on the other hand, have the potential to form a complementary combination of knowledge leading to more impactful research output. The analysis and results highlight the important role of research topic specialization and provide insights for the university- or department-level strategy for research collaborations.

1. Introduction

Decreasing communication and transportation costs largely enabled more and more collaborations in scientific research. As such, research collaboration is regarded as a major approach for advancing science and technology worldwide, as well as developing the scientific impact of the participants, ranging from individual scientists to national systems [1,2,3]. Consequently, scientific research collaboration has been one of the trending topics in bibliometric and scientometric studies.

Aside from the widely acknowledged fact that collaborative research normally yields higher impacts [4], other benefits can also be expected. For example, research collaboration could contribute to individual scientists’ visibility and recognition [1,5], as well as countries’ research capability and competitiveness [6]. While research collaborations are normally studied as co-authorship networks derived from scientific publications, such networks have shown nontrivial patterns, such as core-periphery structure [7] and community structure [8]. Individual scientists may have frequent collaborators from research groups, and countries may have close collaboration patterners. Some collaborations are thus strong and close, while others may be only occasional. Evidence has been found that a variety of factors are shaping such structure of collaboration closeness, including science, economy, geopolitics, and culture [2,9,10].

Universities are the basic units hosting scientists and accordingly, their performance bears great significance for the development of higher education and science. It has been found that an institution’s collaboration network also plays a critical role in its academic performance [11]. However, there is rather scarce literature on inter-university research collaboration in comparison to that on the scientist level and country level. Existing relevant studies mostly focus on the collaboration between university and industry [12,13]. In particular, while countries have been found to generally specialize in some particular research fields [14], a university with dozens of research staff may have also specializations in different research topics. Taking the discipline of Information Science and Library Science as an example, a university (or its information school) may have its specialty in bibliometric studies, while another may be specialized in health informatics. While such research topic specialization largely represents a university’s knowledge composition, its impact on inter-university research collaboration is still largely unknown, which may closely relate to the quality of research and education of universities.

The present paper analyzes the research publications in the discipline of Information Science and Library Science. Applying community detection techniques in the co-occurrence network of keywords, 10 research topics are identified. The specialization of each university across such research topics is thus quantifiable. Accordingly, we explore the following research questions:

(1)

How does the research topic specialization impact a university’s collaborativeness?

(2)

Does the similar research topic specialization promote the closeness and output quality of the collaborations among universities?

2. Related Work

2.1. Research Collaboration

Research collaboration has long been studied via the co-authorship networks, especially on the individual scientist level and country level. Normally, if a research publication has multiple authors, these authors are considered in collaboration with each other; if there are affiliations from multiple countries involved, these countries are considered to be collaborating. Meanwhile, the weights of each collaboration can also be quantified according to how often the participants co-publish with each other. Collaboration networks derived from such simple mechanisms show complex patterns as well as great significance for the development of science and technology. As such, research collaboration has attracted widespread attention from multiple disciplines, especially bibliometric and scientometric studies [15,16] and network science [17,18].

Individual scientists collaborate with each other to contribute their respective insights, knowledge, and sometimes special equipment or materials. As a consequence, collaborative research is normally of higher impact. In addition to the joint input into research, scientists may also exchange knowledge and skills through collaboration. As such, collaborating with others has been an effective approach to improving one’s research impact and career development [2,5]. A scientist’s collaborator network has been found to largely determine his/her future academic impact [19] and productivity [20]. Moreover, the composition of a research team may also significantly influence the research output. It was reported that large teams tend to develop existing research streams, while small teams tend to branch out to new streams [18]. The moderate diversity in the team members’ cultural backgrounds is able to promote the team’s research performance [21]. The new collaborators in a research team are also found able to bring in new knowledge leading to research with higher originalities [22].

International collaborations among countries have been another heated focus in the literature. The collaboration network is found to be organized around a few dominant countries [23], forming a core-periphery structure [7]. Considering the intensity of collaborations, clusters of countries that collaborate closely with each other have emerged. For example, the countries from west Europe and North America are argued to be a cluster, while north Europe is another cluster [24]. Zhu argues that the north European countries, Commonwealth countries, and trans-Pacific countries are three dominant clusters in Communication research [8]. Given such significant regularities, the driving factors of scientific collaborations among countries have been widely discussed. Geographical proximity is normally regarded as the most consistent factor that countries tend to collaborate more closely with others nearby. In addition, economic factors such as the trade relation [25], capability factors such as intellectual and scientific excellence [10], cultural factors such as the common language [9], and political factors such as the co-membership in IGOs [2] have also been found shaping the intensity of research collaborations among countries.

The research collaboration on both the scientist level and country level has attracted widespread attention. The collaboration between scientists is a rather micro perspective, while that between countries is an integrated macro perspective. As a meso-level perspective of the research collaboration, the performance of and collaboration among universities are less-explored in comparison to that among scientists or countries. In the following subsection, we discuss the related work on such meso-level research collaboration.

2.2. Universities in Research Collaboration

Universities play an important role in the development of science and technology since they provide hosts to researchers. Particularly, many research projects and funding are largely supported by one or more universities. In addition, it is primarily in universities where education is linked with scientific research. Accordingly, the performance and ranking of universities are of widespread interest [26,27]. Since scientific research is more and more collaborative, research collaboration has also become a key driver for university development.

The focus of relevant studies lies in the university–industry (UI) collaboration [28], which is found beneficial for both universities [11] and industries [29]. On one hand, industries tend to persistently collaborate with universities, and such persistency is reported to associate with openness to innovation [30]. On the other hand, the tendency of universities to collaborate with industries, such as collaborative research, contract research, consulting, etc., is largely driven by the faculty quality and varies for different disciplines [31]. For example, the technology-oriented disciplines and medical and biological sciences are more collaborative. Close UI collaboration enables the generation of tangible outcomes, which in turn further promote the closeness of the collaboration [32]. In addition, inter-organizational trust contributes to UI collaboration by easing the barriers [33]. The complementarity between the university and industry is also argued to promote UI collaborations, leading to a “honeymoon period” for the start-up stage, while the dissimilarities may disrupt such collaborations [34]. Since UI collaborations are critical for the advancement of science and technology, as well as the economy, governments are keen to promote UI collaborations. As such, the triad of university–industry–government relationship has been widely characterized as the Triple Helix structure [35]. Such trilateral interactions are reported to have a significant impact on the development of research institutes’ academic performances [13], and also on the productibility of industries [36].

While UI collaboration is important for the transformation from knowledge to social and economic impact, inter-university collaboration is equally crucial for the generation of new knowledge. However, there is scarce literature to explore what factors promote the intensity and success of inter-university collaborations. This paper fills in such a gap by exploring the role of research topic specialization of universities on the closeness and impact of inter-university collaborations.

3. Materials and Methods

3.1. Bibliometric Data of Information Science and Library Science

Inspired by rich studies on international and inter-scientist collaborations, we adopt a bibliometric approach to explore the collaborations among universities as reflected in joint scientific publications. The bibliometric data to be analyzed in the present study were downloaded from the Web of Science (WOS), one of the most used databases for scientific publications. The WOS category of “Information Science and Library Science” was chosen as the focal discipline.

The study aims to explore the role of research topic specialization of universities in their collaborations. However, such specialization could be evolving due to two reasons. First, a university’s focus and personnel may change over time, which leads to a change in its specialization. Second, the discipline may also evolve with new topics and new theories emerging. As such, we only consider the publications in a five-year time window, that is, from 2016 to 2020. In the considered short time window, we believe the change of university personnel and discipline could be, to a moderate degree, ignored. Accordingly, the key information of all publications subjected to the category of Information Science and Library Science during the named period is downloaded, including author keywords (WOS code: DE), addresses (WOS code: C1), number of citations (WOS code: TC), year of publication (WOS code: PY), and so on. Removing publications with no author keywords, the final data for analysis consist of 15,431 papers published from 2016 to 2020 in the category of Information Science and Library Science.

For each collected publication, the address information of each author is provided. Based on such information, 1386 affiliations were identified which have at least 5 publication records in the data. Though a significant number of publications have detailed departmental information, we extracted only the first level of the affiliation. For example, from the address “Wuhan Univ, Informat Management Sch, Wuhan 430072, Peoples R China”, we extract the affiliation as Wuhan University. In addition, different campuses of a university were regarded as independent affiliations, such as the University of Wisconsin-Madison and the University of Wisconsin-Milwaukee. Among the identified affiliations, 1156 were universities. The remaining affiliations were mostly independent schools, institutions, and colleges. Meanwhile, there were also a few hospitals and companies. Given the research objectives, we did not consider the hospitals and companies in the analysis, and we refer to the remaining affiliations as “universities” henceforth.

The authors of each publication could come from different affiliations. If a publication involves two different universities (either a single author affiliated to multiple universities, or multiple authors from different universities), it was then considered an inter-university collaborative article. In many articles, different schools of the same university may be presented as different addresses. For articles that involve only one university, regardless of the number of authors or schools/departments from the university, we considered them single-affiliation articles.

3.2. Research Topic Identification via Keyword Network

Research topics, largely reflecting the research interests and specialties of knowledge, are an important factor for both individual scientists and schools/universities. Identification of research topics by analyzing keyword networks is a well-developed approach in bibliometric studies and has been widely applied.

A keyword network is normally constructed by examining the co-occurrence of keywords in publications. The dynamics and structure of knowledge in a discipline or a given set of publications can be thereby investigated. Since the links between keywords (or knowledge) represent a close association, network community detection algorithms can be applied to discover dense clusters of knowledge, which can be further summarized as research topics. Such an approach has been widely implemented to study the evolution and trends of research streams.

In the present study, we focused on the author keywords of each considered article. Applying a Python-based package entitled “pattern”, we firstly transformed all the plural keywords into singular forms. Over 9000 keywords were identified, among which, social media, bibliometric, academic library, and knowledge management were the most frequent. For any pair of keywords, $i$ and $j$, if they appeared simultaneously in the same articles, a link was established connecting them. The weight of each link was calculated as $w_{ij}=n_{ij}/\sqrt{f_i{f}_j}$, where $n_{ij}$ is the number of times the two keywords co-occurred in the same article, and $f_i$ and $f_j$ are the frequencies of the two keywords, respectively. The larger weight of the link indicates the closer association between the knowledge. With keywords as nodes and co-occurrences as links, a keyword network was generated as shown in Figure 1, where node sizes are proportional to the logarithm of keyword frequencies and link widths are proportional to the co-occurrence frequencies. Only the links with weights larger than 0.02 are shown in the network.

While a number of community detection algorithms are available, we applied a popular algorithm entitled Louvain to the constructed keyword network. In the community detection, not only the linkage structure among the keywords but also the weights of links were considered. The Louvain algorithm was set to partition all the keywords into 20 clusters, which were further reduced to 10 topics for the following reasons. Firstly, there were clusters with only a handful of keywords, which were removed if the keywords were not frequent or merged into another related cluster. Secondly, some clusters may be closely associated with each other and thus were combined. According to the keywords, with consultancy from experts in the discipline, we summarized the topic of each cluster as shown in

Table 1. Results of topic modeling.

Topic	Keywords	Frequency	Frequent Keywords
Library and Archive	1359	10,643	Academic Library; Information Literacy; Case Study; Library; Information Retrieval
Bibliometrics	1264	10,417	Bibliometric; Citation Analysis; Collaboration; Open Access; Altmetric; Scientometric
Health Informatics	1181	11,090	Qualitative; Electronic Health Record; Communication; Healthcare; Content Analysis
Information System	1054	6683	Information System; Trust; Adoption; Usability; Developing Country; Continuance
Social Media	1009	7245	Social Media; Twitter; Social Network; Facebook; Internet; Media; Social Networking Site
Information Technology	1008	5780	Big Data; ICT; Innovation; Information Technology; Digital Divide; Development; Privacy
Knowledge Management	724	4478	Knowledge Management; Knowledge Sharing; Social Capital; Knowledge Creation
Analytical Methods	812	3934	Machine Learning; Text Mining; Natural Language Processing; Deep Learning
Marketing	386	1751	Online Review; Recommender System; UGC; Electronic Commerce; Marketing; EWOM
E-Government	277	1677	E-Government; Transparency; Smart City; Governance; Regulation; Accountability

. The keyword network shown in Figure 1 is not a random sample but includes only the 20 most frequent keywords from each topic.

To summarize, 10 topics were identified by applying the Louvain algorithm on the keyword network, including Library and Archive, Bibliometrics, Health Informatics, Information System, Social Media, Information Technology, Knowledge Management, Analytical Methods, Marketing, and E-Government. The topics were defined according to keywords, rather than articles. In other words, a particular article could involve keywords from multiple topics. For example, an article may apply text mining techniques to study the altimetric problems, and thus contains keywords from both topics of Bibliometrics and Analytical Methods.

4. Results

To explore the impact of research topic specialization on inter-university collaboration, we carried out analysis from three different levels, namely the article level, university level, and dyadic collaboration level.

4.1. Collaborative Articles across Research Topics

As discussed earlier, each article could have keywords from multiple topics. Accordingly, each article $a$ can be represented by a vector of length of 10, denoted as $N^a=\left\{n_1^a,n_2^a,\cdots ,n_{10}^a\right\}$, where $n_t^a$ is the number of keywords in article $a$ that are subject to topic $t$. Here we explore the impact of such a vector on the collaborativeness of each article, which can answer a research question that which research topic attracts more cross-university collaborations. To do so, we analyze a logistic regression model, i.e., $ColAr{t}_a=\beta N^a+ϵ$, where $ColAr{t}_a=1$ if the article $a$ is collaborated by multiple universities and $ColAr{t}_a=0$ otherwise.

As reported in

Table 2. Collaborativeness of research topics.

	Estimates	Std. Err.
Library and Archive	−0.213 ***	0.015
Bibliometrics	−0.049 ***	0.014
Health Informatics	0.051 ***	0.010
Information System	0.132 ***	0.020
Social Media	−0.098 ***	0.015
Information Technology	−0.031	0.020
Knowledge Management	0.062 ***	0.023
Analytical Methods	0.050 **	0.025
Marketing	0.200 ***	0.040
E-Government	0.010	0.035
Constant	0.395 ***	0.039
Observations	15,431
R-Squared	0.031

** p < 0.01; *** p < 0.001.

, articles with more keywords in topics of Library and Archive ($\beta =-0.213, p<0.001$), Bibliometrics ($\beta =-0.049, p<0.001$), and Social Media ($\beta =-0.098,p<0.001$), are less likely to be collaborative with multiple universities. In contrast, articles with more keywords on Health Informatics ($\beta =0.051,p<0.001$), Information System ($\beta =0.132,p<0.001$), Knowledge Management ($\beta =0.062,p<0.001$), Analytical Methods ($\beta =0.050,p<0.01$), and Marketing ($\beta =0.200,p<0.001$) tend to be published via cross-university collaborations. Such results indicate that the research on topics of Health Informatics, Information systems, Knowledge Management, Analytical Methods, and Marketing normally requires inter-university collaborations, while the research on Library and Archive, Bibliometrics, and Social Media involves fewer collaborations.

4.2. Research Topic Diversity and Its Impact on University Collaborativeness

At the university level, we explore the diversity of each university’s research topic specialization. Due to the limited research staff in a discipline (in this paper, Information Science and Library Science), the research of these research staff is unlikely to evenly cover all the topics. A university may thus have a particular specialization in different research topics. To avoid confounding effects, here we defined such research topic specialization of a university according to only the single-affiliation articles, i.e., the articles authored by only the target university.

For a university $u$, suppose its single-affiliation articles have $W$ keywords in total, including those repetitive keywords. If there are $w_t$ keywords that are subject to the topic $t$, we define the university $u$’s specialization in topic $t$ as $p_t^u=w_t/W$. Accordingly, the research topic specialization of the university $u$ can also be characterized as a vector $P^u=\left\{p_1^u,p_2^u,\cdots ,p_{10}^u\right\}$, and we have ${\displaystyle \sum }_{t=1}^{10}{p}_t^u=1$ for any university. Figure 2a–c shows three examples of research topic specialization. UC Berkeley is strongly specialized in Library and Archive; about 68% of keywords in its single-affiliation articles are subject to such a topic. In addition to this major topic, UC Berkeley has minor specializations in topics of Health Informatics, Information System, Social Media, Information Technology, and E-Government, but has no keywords in topics of Bibliometric, Knowledge Management, Analytical Methods, and Marketing. In contrast, the University of Sheffield and Nanjing University have keywords in each of the 10 topics. While the University of Sheffield specializes slightly more in Library and Archive and Health Informatics, Nanjing university specializes more in Bibliometric and Marketing.

Research topic specialization largely reflects a university’s focus. Here, we further defined the diversity of research topic specialization according to Shannon entropy, which can be calculated as follows.

$$di{v}_u={\displaystyle \sum }_{t=1}^{10}-p_t^u{\log }_2{p}_t^u.$$

(1)

Such a value describes the extent to which the research topic specialization is evenly distributed across different research topics. A large value of diversity indicates the even distribution, while a low value indicates centralized distribution on some particular topics. For example, since UC Berkeley has a very strong specialization in Library and Archive, it has a research topic diversity of $div=1.61$. On the other hand, the specialization distributions of the University of Sheffield and Nanjing University are more even, thus the two universities have diversities of 2.87 and 2.84, respectively, which are among the highest values. Figure 2d reports the distribution of research topic diversities of 376 universities that have published at least 5 single-affiliation articles. It is shown that most universities have topic diversities ranging from 1.5 to 3.0.

The impact of research topic diversity on the collaborativeness of universities was further analyzed via a linear regression model. The collaborativeness of a university, denoted as $co{l}_u$, was calculated as the fraction of collaborative articles over all articles published by the university, ranging from 0 to 1. Several control variables were also considered. We collected the QS ranking (top-50, 2020 list) of the subject “Library and Information Management”, and construct a variable $Q{S}_u$ to indicate whether the university is QS-ranked ($Q{S}_u=1$) or not ($Q{S}_u=0$). The number of single-affiliation articles, denoted as $NS{A}_u$, was also considered as a control variable. The impact of an article is defined as its citations divided by the mean value of citations over all publications from the same year, to remove the effects of publication time. Thus, the research impact of a university $impac{t}_u$ is thus the average impact of all its single-affiliation articles. Accordingly, we analyzed the following model.

$$co{l}_u={\beta }_{1}di{v}_u+{\beta }_{2}Q{S}_u+{\beta }_3\log \left(NS{A}_u\right)+{\beta }_{4}impac{t}_u+{\beta }_0.$$

(2)

As shown in

Table 3. Collaborativeness of universities.

	Estimates	Std. Err.
Topic diversity	0.085 ***	0.020
QS ranked (top-50)	0.108 ***	0.032
Single-affiliation articles	−0.049 **	0.021
Research impact	0.009	0.011
Constant	0.526 ***	0.058
Observations	376
R-Squared	0.074

** p < 0.01; *** p < 0.001.

, topic diversity has a significant positive impact on the collaborativeness of universities ($\beta =0.085,p<0.001$). Thus, the universities with diverse research topics have a stronger tendency to collaborate with other universities in joint publications. Meanwhile, the top universities as ranked by QS also tend to have more cross-university collaborations ($\beta =0.108,p<0.001$). In addition, the more single-affiliation articles a university published, the less it would collaborate with other universities ($\beta =-0.049,p<0.01$). On the other hand, the research impact of a university does not have a significant influence on its collaborativeness.

4.3. Impact of Research Topic Similarity on Inter-University Collaboration

An inter-university collaboration network can be constructed by examining how often each pair of universities collaborate with each other. Suppose two universities $i$ and $j$ have jointly published $c_{ij}$ articles. The larger values of $c_{ij}$ indicates more joint publications. However, there are some universities that have published a lot of articles, and thus would naturally have more collaborative articles with a particular university. To avoid possible bias, we adopted a Salton index to calculate the collaboration closeness between university pairs as follows.

$$closenes{s}_{ij}=\frac{c_{ij}}{\sqrt{c_i\cdot c_j}}.$$

(3)

In the equation, $c_i$ and $c_j$ are the number of collaborative articles published by university $i$ and $j$, respectively. Such a metric $closenes{s}_{ij}$ describes how much of the universities’ collaborative articles involve each other, and thus a larger value indicates closer collaboration. Figure 3 shows the collaboration network among universities ranked in the QS subject “Library and Information Management”. While there are 50 universities ranked, the illustrated network includes only 43 of them since we applied a criterion that only universities with at least 5 single-affiliation articles be analyzed (for the purpose of calculating research topic specialization). In the network, node size is proportional to the logarithm of the number of single-affiliation articles, while the edge width is proportional to the collaboration closeness as defined in Equation (3). Only the collaboration links with closeness larger than 0.02 are shown.

Aside from the closeness of inter-university collaborations, we also analyzed the impact of such collaborations. As defined earlier, the impact of an article is the normalized citations, which is the raw citations divided by the mean citations of the corresponding publication year. The impact of an inter-university collaboration, denoted by $ColImpac{t}_{ij}$, is thus defined as the average impact of the articles collaborated on by the two universities.

While each university has its own research topic specialization, the similarity of such specialization of two universities $i$ and $j$ can be calculated according to a Pearson correlation index, as follows:

$$si{m}_{ij}=Pearson\left(P^i,P^j\right)=\frac{{{\displaystyle \sum }}_{t=1}^{10}\left(p_t^i-0.1\right)\left(p_t^j-0.1\right)}{\sqrt{{{\displaystyle \sum }}_{t=1}^{10}{\left(p_t^i-0.1\right)}^2}\sqrt{{{\displaystyle \sum }}_{t=1}^{10}{\left(p_t^j-0.1\right)}^2}}$$

(4)

It needs to be noted that, since each specialization vector $P$ contains 10 values with a summation of 1, the mean of each value is always 0.1 for the specialization vector of any university. The research topic similarity between the two universities ranges from −1 to 1, with negative values representing opposite specializations while positive values representing similar specializations. Taking the three universities shown in Figure 2a–c as examples, the University of Sheffield and UC Berkeley have a research topic similarity of 0.645, due to the same specialization in the topic of Library and Archive. On the other hand, the University of Sheffield prefers Library and Archive, and Health Informatics, while Nanjing University prefers Bibliometric and Marketing, and thus they have a topic similarity of only 0.131. For collaborations among all universities with at least 5 single-affiliation articles, the average research topic similarity takes a value of 0.366. Such an average value is not 0 due to the uneven distribution of research topics in total. For example, Bibliometric and Health Informatics are naturally more popular than Marketing and E-Government in the discipline. In the illustrated network shown in Figure 3, we classified the collaboration links among the top universities into two types. The red links represent collaborations that the two universities that have topic similarities larger than the average value of 0.366, while the black links are those with similarities less than the average. It is apparent that most of the collaborations happen between top universities with similar research topic specializations. This is partially a piece of evidence showing the positive impact of research topic similarity on close inter-university collaborations.

To further quantify the influence of research topic similarity between two universities on the closeness and impact of their research collaborations, we applied the following two economic models.

$$closenes{s}_{ij}={\beta }_{1}si{m}_{ij}+{\beta }_2\log \left(SumNS{A}_{ij}\right)+{\beta }_3\log \left(DifNS{A}_{ij}\right)+{\beta }_{4}SumIm{p}_{ij}+{\beta }_{5}DifIm{p}_{ij}+{\beta }_0$$

(5)

$$ColImpac{t}_{ij}={\beta }_{1}si{m}_{ij}+{\beta }_2\log \left(SumNS{A}_{ij}\right)+{\beta }_3\log \left(DifNS{A}_{ij}\right)+{\beta }_{4}SumIm{p}_{ij}+{\beta }_{5}DifIm{p}_{ij}+{\beta }_0$$

(6)

In the equations, $SumNS{A}_{ij}=NS{A}_i+NS{A}_j$ is the summation of the two universities’ single-affiliation articles, while $DifNS{A}_{ij}=\left|NS{A}_i-NS{A}_j\right|$ is the difference in such values. Similarly, $SumIm{p}_{ij}=impac{t}_i+impac{t}_j$ and $DifIm{p}_{ij}=\left|impac{t}_i-impac{t}_j\right|$ are the summation of and difference in the two universities’ average impact of single-affiliation articles.

The unit of analysis is collaboration, which is dyadic and networked. Potential inter-correlations may exist in such kinds of data which may lead to the inaccurate estimation of the significance of variables. Here we applied the Quadratic Assignment Procedure (QAP) to avoid such an issue. The QAP shuffles the networked data records for a given number of times (in this paper, 100 times) and compares the empirical data with such random data. As such, the $p$-values are estimated more accurately from the comparison. Applying QAP to the above-introduced models, the results are reported in

Table 4. The closeness and impact of inter-university collaborations.

	Collaboration Closeness	Collaboration Impact
Topic similarity	0.791 ***	−0.082 ***
Summation of single-affiliation articles	1.718 ***	0.361 ***
Difference in single-affiliation articles	−0.862 ***	−0.128 ***
Summation of average impact	0.177 ***	0.119 ***
Difference in average impact	0.365 ***	−0.084 ***
Constant	0.014	0.001
Observations	5866	5866
R-Squared	0.554	0.402

*** p < 0.001.

.

The respective capabilities of the two universities have a significant influence on the closeness and impact of their collaboration. Summation of single-affiliation articles and summation of average impact show positive effects on both collaboration closeness ($\beta =1.718,p<0.001$ for $SumNS{A}_{ij}$ and $\beta =0.177,p<0.001$ for $SumIm{p}_{ij}$) and collaboration impact ($\beta =0.361,p<0.001$ for $SumNS{A}_{ij}$ and $\beta =0.119,p<0.001$ for $SumIm{p}_{ij}$). It is suggested that close collaborations tend to happen between universities that can publish a lot of research articles on their own and generate high impact. Such collaborated articles are normally also of high impact. On the other hand, if two universities have very different capabilities to publish on their own, i.e., a large value for the difference in single-affiliation articles $DifNS{A}_{ij}$, they tend to collaborate less ($\beta =-0.862,p<0.001$) and such collaborations are of less impact ($\beta =-0.128,p<0.001$). The difference in the two universities’ respective research impact, $DifIm{p}_{ij}$, shows mixed effects for the closeness and impact of collaboration. While a university with low research impact and a university with high research impact tend to have close collaboration ($\beta =0.365,p<0.001$), such collaborations are of low impact ($\beta =-0.084,p<0.001$).

As the focal explanatory variable, the similarity between the two universities’ research topic specializations also shows mixed effects. Collaborations between universities with very similar research topics tend to be closer ($\beta =0.791,p<0.001$), but on the other hand are normally less impactful ($\beta =-0.082,p<0.001$). Such a result indicates that universities tend to collaborate with the ones that have similar research topics, leading to close collaboration. However, the research topic specialization of a university, to a large extent, represents the knowledge structure of the university. Due to the very similar research topics, the universities are actually not complementing each other with novel knowledge, leading to the low impact of the collaborated research output. In contrast, two universities with different research topic specializations could contribute and combine their respective strength in different aspects of knowledge, and thereby have the potential to generate new knowledge with higher impact.

5. Conclusions and Discussion

Universities play a critical role in the development of science and technology, and collaboration research has been a significant driver. While most previous studies focused on the collaborations between universities and industries, the present paper examines the inter-university research collaboration and explores the impact of research topic specializations on it. Taking the discipline of Information Science and Library Science as an example, inter-university collaboration was revealed to be largely determined by the research topic specializations of universities. It was found that universities with diverse research topic specializations have a stronger tendency to collaborate with others. The similar research topic specializations between two universities make their collaboration closer, but at the same time may lower the impact of such collaboration.

The study offers novel implications for both the theoretical understanding of research collaboration, but also the strategic management of research universities. In fact, research collaboration has been intensively investigated at the country level, where the structure and closeness of the collaborations are widely discussed. The significant impact of widespread proximities on international collaboration closeness has been uncovered, including geographical distance [37,38], economic proximity [10,39], cultural similarity [2,9], and so on. The results in this paper on inter-university collaboration are not only in line with the previous observations, but also provide a new dimension of proximity, i.e., the proximity of research topic specialization. Such a new proximity dimension was found to promote the closeness of inter-university collaboration, and thus may also have the potential to drive international collaboration. In addition, previous research has addressed the importance of the complementary capabilities of the two collaborating parties. Estrada et al. argue that the complementarities stimulate the knowledge exchange process in the university-industry collaboration [34]. Individuals are also found to seek collaborators with complementary professional skills to enhance utility [40]. The resource complementarity between enterprises also plays a critical role in collaboration performance [41]. Knowledge, as an important resource in scientific research, is shown in this paper to have similar complementary effects. Although universities tend to collaborate more with others who have similar research topic specializations, such universities may not best complement each other with new knowledge due to the overlapped research topics. This is largely the reason that collaborations between universities with dissimilar research topics are more impactful. The combination of knowledge from different research topics is thus indicated to have the potential to yield new insights. Universities or research groups are advised to seek collaborators with complementary knowledge.

There are still several limitations in the present study that could be addressed in future research. First, only 10 topics were summarized for the entire discipline of Information Science and Library Science, which may not be exactly accurate. For example, a major keyword in the cluster of Health Informatics is “qualitative”. However, such a keyword may appear in research articles on a wide range of topics. Furthermore, the knowledge structure of any discipline could be complex and different topics may nest with each other. The interdisciplinary topics thus should also be analyzed carefully. In addition, the research staff of even a given information school could still publish in various journals from different disciplines. Thus, the data applied in the present study, that is, the publications from only the Information Science and Library Science, cannot fully represent the knowledge specialization of a university. Second, though the different research topic specializations have been shown to be effective in promoting the impact of collaboration, what combinations of different topics are optimal to achieve such effect shall be further explored. Third, we only examined the research topic specializations and collaborations of universities in a static period of time, while such specialization and collaborations have been found to evolve over time [42]. The possible interaction between the inter-university collaboration and research topic specialization over time shall also be studied. How the results observed in the present study would differ with the evolution of research topics in the given discipline also needs to be future explored. Last, while proximities of geographical distance, economic development, language, etc. have been widely addressed for international collaboration, the analysis for inter-university collaboration in this paper does not consider such control variables due to the difficulties of collecting relevant data.