Anatomising the impact of ResearchGate followers and followings on influence identification

Abstract

Influence analysis, derived from Social Network Analysis (SNA), is extremely useful in academic literature analytic. Different Academic Social Network Sites (ASNS) have been widely examined for influence analysis in terms of co-authorship and co-citation networks. The impact of other network-based features, such as followers and followings, provided by ASNS such as ResearchGate (RG) and Academia is yet to be anatomised. As proven in ingrained social theories, the followers and followings have significant impact in influence prorogation. This research aims at examining the same in one of the widely adopted ASNS, RG. The rendering process is developed to render real-time RG information, which is modelled into graph. Standard centrality measures are implemented to identify influential users from the constructed RG graph. Each centrality measure gives a list of top-k influential RG users. The results are compared with RGScore and Total Research Interest (TRI) to discover the most effective centrality measure. Betweenness and closeness centrality measures have shown the outperforming results compared with others. A procedure is established to discover influential RG users that are commonly present in all top-k centrality results to identify dominant skills, affiliations, departments and locations from the rendered data.

Keywords

Academic social network sites centrality measures connected data graph database influence analysis ResearchGate

1. Introduction

The social media platforms provide means of social communication or personal information sharing [1,2]. Likewise, Academic Social Network Sites (ASNS) are profusely used for various research centric activities such as research sharing, academic collaborations, scientific communication, feedback collection and technical information dissemination [3]. ASNS provide a flexible medium to manage the explosive growth in research activities compared with traditional scientific outlets, such as journals and conferences [4]. Prevalent exploitation of ASNS, such as ResearchGate (RG), Google Scholar (GS), Academia.edu and Mendeley, has opened a novel paradigms of research in the field of scientific influence finding, inspired from social influence finding [5,6]. Finding scientific influence in terms of scholars [7,8], research topics [9], journals [10,11] and publications [12 –14] using co-authorship and co-citation networks is a well-known concept, as these networks are available for many platforms such as Mendeley, Academia.edu, Scopus, GS, Digital Bibliography and Library Project (DBLP) and Association for Computing Machinery (ACM).

Apart from the co-authorship and co-citation connections, ASNS, such as RG and Academia, provide user-to-user connections in terms of followers and followings. It is observed that the perspective of follower and following connections is yet to be explored with respect to scientific influence identification. Recently conducted surveys demonstrate the higher utilisation of RG among other well-known ASNS in terms of number of registered profiles [6,15] and active participation [16 –18]. This motivates our research focus mainly on RG. In this research, we explore RG in terms of connected graph with an aim of anatomising the impact of follower and following connections on influence analysis.

Information shared on RG is categorised into user profile content: research items, citations, reads, recommendations, department, position, affiliation, etc., and user-to-user links: followers, followings, co-authors and co-citations. A data rendering process is developed to render RG users’ profile content based on follower and following connections present among users. The rendered information is modelled into graph using graph modelling technique. The proposed methodology is an initial step to generate and utilise graph-based model of RG information to find influence considering user content and user-to-user followers and followings information present on RG. The visual graph layout helps in identifying how the influence propagates through follower and following connections.

Standard graph centrality measures – betweenness centrality (BC), closeness centrality (CC), eigenvector centrality (EvC) and PageRank centrality (PRC) – are implemented on constructed RG graph for influence identification. The concept is to leverage the follower and following connections to see how the influence of one RG user affects the influence score of other connected RG users through centrality measures. Each centrality measure identifies a list of top-k influential users from RG graph. The results are compared in terms of ranking correlations with influence assessment scores – RGScore and Total Research Interest (TRI) – given to each registered and active user by RG. RGScore is mentioned to be calculated based on the research in author’s RG profile and the interactions of other authors with it. TRI is mentioned as the sum of the research interest of each research item present in authors’ RG profile.

Also, a process is established to find the common influence, in which a list having distinct influential users that are present for at least once in centrality top-k lists is computed. The significant characteristics – skills, affiliations, departments and locations of identified common influential users – are discovered.

In section 2, the pioneer work in the domain of influence finding in ASNS is briefly reviewed. In section 3, the methodological framework is discussed. In section 4, the results are presented along with the detailed discussion and significant reasoning. Section 5 summarises this research with future expansion possibilities. The sample XML graph model of the prepared RG graph, the list of abbreviations and the list of notations are presented in Appendix 1.

2. ASNS and influence

The web-based ASNS are serving baseline mediums in achieving the competitive success in research sector [19]. In research community contextual and influence finding paradigms, ASNS are receiving intense attention. They offer broadly adaptable, efficaciously available and digitised medium to the prominent researchers of diverse disciplines and geologically scatter areas to share and access ad hoc, pre-print or full-fledged research and to collaborate, communicate and acquire assistance from scientific experts [1,3]. For such reasons, ASNS such as RG, Mendeley or Academia.edu claim to have several millions of recorded users [4,5]. Massive scientific data in relevant fields of research can provide pivotal information for sound scientific research and analysis; although, the present literature discloses a little interest in finding influential users/researchers based on their real-time research activities on any ASNS platform.

Each user in ASNS can possibly extend his scientific impact; however, a few users demonstrate their supremacy in scientific impact propagation [20,21]. Scientific influence finding is recognising these predominant (influential) users who can affect the assessments of different users by viral advancements or explicit activities. The recognition of such influential users is valuable for showcasing, promoting or spreading influence [22]. Identification of research domains, locations, department of work and so on of these influential users provides bright possibilities of communication and collaboration for emerging researchers [23]. The identified potential users with their specific domain of research interest can be invited for various research activities [24,25] such as article review, expert talk, conference speaker, multi-domain research collaboration and more [4,26].

2.1. Graph-based influence analysis in ASNS

Modelling social media information as a graph of connected data has its roots in many profound applications such as influence finding, topic analysis, clustering, expert finding, community detection and recommendation systems [22]. Graphs are demonstrated to be profoundly proficient for different social media data analytics [27 –32]; among which, influence finding is preeminent.

Various ingrained social theories illustrate that the impact (influence) of one entity propagates to other connected entity and such influence can be effectively analysed using graph models [27]. For social media platforms such as Twitter and Facebook, the follower and following relations are proven to be substantial while analysing influence propagation and identifying influential entities. Furthermore, the technological advancements provide various competent mechanisms for graph storage and graph information retrieval [30]. Graph-based centrality measures [33] are widely employed for influence analysis in many application domains such as social media [34 –37], web platforms [37 –39], finance [40] and academic literature [41,42].

In academic literature domain, pioneer of work has been conducted in the field of finding influential researchers/scholars. Majority of such work utilises the available co-authorship [25,43 –50] and co-citation networks [7,10 –13,49,51,52] from graph data repositories. Apart from using the historic information of co-authorship and co-citation, identification of real-time influence by fetching the real-time linked data serves an establishment of developing potential research in academic literature analytic [53].

2.2. Related work

This research focuses mainly on analysing the impact of RG followers and followings on influence identification. The related work is discussed with respect to the existing studies that (1) compare RG with other well-known ASNS and (2) conduct influence analytics in RG.

2.2.1. Comparing RG and other ASNS

In literature, various analysis and surveys are conducted on well-known ASNS to identify their popularity in terms of total registered profiles, usage frequency and degree of activeness.

An analysis was conducted on the distribution of Spanish National Research Council researchers’ profiles on Academia.edu, GS and RG [15]. The study discovered the higher utilisation of RG in terms of number of registered profiles with the statistics of 4001, 2036 and 1156 profiles on RG, GS and Academia.edu, respectively. A study offering an overview of established and emerging ASNS – Scopus, WoS, PubMed, RG, GS, Academia.edu, Open Researcher and Contributor Identification (ORCID) and Publons – disclosed that RG is a widely utilised platform [18]. Profiles of 400 RG users were examined with respect to RG and h-index (GS and Scopus) [54]. RG was identified to be a leading ASNS in a way it employs researchers’ data.

A comparative overview of various ASNS, such as RG, GS, ORCID, Scinapse, Semantic Scholar, ScienceOpen and Kudos, disclosed RG to be superior among others due to its prominent status features [55]. A survey on Academia.edu, Mendeley, RG, MyScienceWork, Humanities Common, Social Science research Network, Profology and Trellis was carried out to learn their usage frequency [17]. The results showed substantially more frequent usage of RG among all with 85% of users using RG least monthly. Another survey conducted on Academia.edu, Mendeley and RG to measure the degree of activeness revealed RG receiving the greatest attention [16]. An assessment of researchers’ presence was performed with respect to four platforms: ORCID, ResearcherID, Academia.edu and RG [56]. The analysis disclosed that out of 1047 researchers studied, RG had by far the highest number of researchers (54.3%) present.

2.2.2. Influence analytics in RG

RG facilitates researchers to generate their profiles, share their research [26], discover the peers with expertise in relevant domains [20] and get insights into inclining research [23,24]. In existing studies, majorly the profile features present on RG are analysed statistically to identify the influence with different prospects. Few studies also explore the co-author networks of RG to identify influential RG users.

A study was conducted to explore the real-time RG publications from Canadian computer science researchers to analyse the correlations among diverse profile features of RG including views, publications, downloads, citations, questions, answers and followers [57]. The research investigations showed that the Canadian researchers are highly active in research associations and scientific information sharing while the statistics illustrated the moderate correlations among various RG profile features such as number of views, downloads and citations. The number of views was around 2 (1/2) times the number of downloads and the number of citations was around 2/3 times the number of downloads for the collected data. An analysis was carried out to inspect the effect of RG statistics on institute ranking [58], in which, the results unveiled moderate correlation.

RG network of Macedonian University was explored for specific research areas [59]. Centrality measures were used to find the influential researchers with multi-domain knowledge. A cluster analysis was performed on RG based on various RG profile features [60]. Three clusters – Active users, Representers and Lurkers – were created. The results disclosed that in the collected data, RG users’ distribution in respective clusters was 8%, 4% and 88%. An analysis was conducted on RG users from 61 US research universities at different research activity stages (categorised by the Carnegie Classification of Institutions of Higher Education) to inspect the impact of institutional differences on RG reputation metrics [4]. The results confirmed that RG is a research-oriented academic social networking site that thoroughly and convincingly mirrors the research activity level of institutions. RG helps in maximising the influence for emerging researchers by increasing RGScores, publications visibility, citations, profile views and followers.

Sample of 4800 RG profiles was analysed to identify the communication patterns and demographic attributes [61]. The influence of age, status and institutional rankings was measured against the researchers’ network activities. The outcome suggested that age or status had no influence on network activities of researchers, whereas the institutional rankings demonstrated the influence on network behaviour. A study to examine the interrelations of scientists’ network communications in RG and their real-life proficient accomplishments was performed in [62]. The results demonstrated that the researchers who were not actively involved in RG network activities had little privilege on the role of academic influential person.

A study was conducted using network science to create RG collaboration network of Iranian Scientific Institutions [63]. The results demonstrated that geographic location closeness and ethnic attributes have influential roles in academic collaboration network establishment. Also, the popular scientific centres in the capital city of Iran, Tehran, have a significant influence on the production flow of scientific activities. A sample of 77,902 RG users from 61 US research universities at different research activity levels was analysed [64]. The sample users were categorised into six groups based on their affiliated departmental disciplines as stated on their RG profiles. The results showed that user participation and user characteristics vary by discipline. In addition, users from higher research activity level universities are identified to be influential in RG metrics regardless of discipline. A study investigated the characteristic usage behaviours of Essential Science Indicator (ESI) highly cited researchers on RG [65]. The results showed that 45.21% of ESI researchers from diverse disciplines have registered with RG. The Spearman correlation denoted that information sharing behaviours have a strong correlation with researchers’ influence compared with social interaction behaviours.

3. Methodological framework

The recent surveys demonstrate RG as greatly utilised ASNS among others, but it is yet to be explored with respect to influence analysis in terms of followers and followings present on RG using real-time RG information.

The significant contributions of this research are as follows:

Render real-time linked information of RG based on follower and following connections

Formulate collected linked information into a graph model and implement graph-based centrality measures on the constructed RG graph

Measure the relevance of obtained results with respect to RGScore and TRI

Develop a method for collective influence finding and explore the properties of commonly present influential RG users in centrality-based top-k results

The proposed methodological framework for anatomising the impact of RG followers and followings constitutes five sub procedures: data rendering, graph modelling, graph construction, implementation of centrality measures and collective influence finding as demonstrated in Figure 1.

Figure 1.

The proposed methodological framework.

3.1. Data rendering

The data rendering process contemplates following qualitative observations.

3.1.1. Observation 1

For the known influencing researcher R in RG having F number of followers and G number of followings, let the influence score of f in F and g in G be I_scf and I_scg, respectively. For $f \in F$ , where f > F/2 and $g \in G$ , where g > G/2; Probability PF states Iscf < Iscg.

3.1.2. Observation 2

In a given time span t_p, unlike social media networks, research networks tend to undergo a smaller number of changes in terms of content and links. Henceforth, this research does not include the concept of evolving network. It rather takes into account the real-time RG information at a specific time t.

While for many ASNS, the co-authorship and co-citation graphs are available, it is challenging to collect real-time linked RG information. Unlike social media platforms such as Twitter, RG does not provide any Application Programming Interface (API) but the structured organisation of information present on RG can be rendered using the ‘web rendering’ concept. A rendering process, specifically designed and accurately correlated with the structure of RG, is developed. It renders publicly accessible user content and user-to-user associations. Here, user content and user demographics, user-to-user links and associations words are used interchangeably.

The rendering process has two sub functions: ContentRender and LinkRender. To demonstrate the working of both the functions, consider a sample RG network as displayed in Figure 2. For any selected target user, U, U₁, U₂ and U₃ are the first level of followings (displayed as FW₁). Based on the Observation 1, for a selected target user U, only the followings are rendered in first phase. For the first level of followings of U, the followers (depicted as FL₂) and followings (denoted as FW₂) are rendered in second phase. The follower and following links are rendered using LinkRender, whereas for every user including U, the profile content is rendered using ContentRender.

Figure 2.

Sample RG network.

The algorithmic steps of the developed data rendering process are demonstrated in Algorithm 1. Line number 3–10 demonstrate the working of ContentRender, while line number 11–18 depict the functioning of LinkRender. The rendered content is stored in file C and the rendered associations are stored in file R.

Algorithm 1.

Data rendering.

Input: Initial User Profile U
Output: Content File C and Relation File R
1: C

\leftarrow

ContentRender(U) ▹Render properties of U
2: R

\leftarrow

LinkRender(U) ▹Render associations of U
3: FunctionContentRender(U):
4: i

\leftarrow

1
5: fori = 1 to n do ▹ Where n is total no. of users profiles to be rendered
6: forj = 1 to p do
7: Render P[i] ▹ Where p is total no. properties to be rendered from one user profile
8: end for
9: end for
10: End Function
11: FunctionLinkRender(U):
12: i

\leftarrow

1
13: Render FW₁▹ FW₁ = FW₁₁, FW₁₂,…., FW_1w ▹ Set of First Level Followings of U
14: fori = 1 to FW_1w do
15: Render FL₂▹ FL₂ = FL₂₁, FL₂₂,…., FL₂₁ ▹ Set of Second Level Followers of U
16: Render FW₂▹ FW₂ = FW₂₁, FW₂₂,…., FW_2w ▹ Set of Second Level Followings of U
17: end for
18: End Function

FW: Following of Node N; FL: Follower of Node N.

3.2. Graph modelling

To transform the rendered RG information into graph structure, a graph model is prepared. Employing the prepared graph model presented in Algorithm 2, the collected information is transformed into nodes, relations and properties forming a connected RG graph.

Algorithm 2.

Graph modelling.

Input: Content File C and Relation File R
Output: Set of Nodes N = N₁, N₂,…., N_n
Set of Relations R = R₁, R₂,…., R_r
Set of Properties P = P₁, P₂,…, P_p
1: fori = 1 to n do
2: Create Node N_i ▹Where n is total no. of users in Content File C
3: Set N_i.Properties for P = P₁to P_p
4: end for
5: fori = 1 to n do ▹Where n = FW_lw rendered by LinkRender(U)
6: Set N_i.Relations for R = R₁ to R_r
7: end for

As displayed in Algorithm 2, line number 1–4 depict that for each content row in input file C, the nodes are modelled. Each rendered property is attached with the respective node. Line number 5–7 denote that for each row in input file R, the relations are modelled among nodes.

3.3. Graph construction

The set of nodes, relations and properties are imported into Neo4j - the leading graph database - to construct RG graph. Cypher Query Language (CQL) is used to interact with the graph in Neo4j. The pseudo-code of the graph construction process is demonstrated in Algorithm 3. Line number 1 displays the initialisation of Neo4j, line number 2–4, respectively, show the import process of nodes, properties and relations in Neo4j.

Algorithm 3.

Graph construction.

Input: Set of Nodes N = N₁, N₂,…., N_n
Set of Relations R = R₁, R₂,…., R_r
Set of Properties P = P₁, P₂,…, P_p
Output : G (N,R)
1: Step 1: Initiate Neo4j instance ▹

\forall n \in N

\forall p \in P

\forall r \in R

2: Step 2: CQL_load(n)
3: Step 3: CQL_create(p)
4: Step 4: CQL_create(r)

N: number of nodes; R: number of relations.

Each ‘node’ in the RG graph represents an RG user while their associations are modelled as ‘directed edges’ with label ‘Following’ and ‘Follower’. N nodes are created and R relations are formed among nodes based on rendered follower and following information.

3.4. Centrality measures

To grasp a comprehensive impression on rendered RG information, we conducted a study to determine which users are at the ‘centre’ of the whole graph; therefore, graph centrality measures are implemented.

The notion of finding influential nodes is subjective depending on how the ‘influence’ is defined. Different centrality measures use different metrics which define the importance of a node from different perspectives. More specifically in this research, four standard centrality measures – BC, CC, EvC and PRC – are adopted and implemented in iterative manner on the graph to check their performance in finding influential users. Each centrality measure computes the list of top-k influential users as mentioned in Algorithm 4. Line number 1–4, respectively, demonstrate the calculation of BC, CC, EvC and PRC for each node in the constructed RG graph. The obtained results are discussed in section ‘Experimental analysis and results’. The evaluation of implemented centrality measures is performed in terms of ranking correlations with respect to RGScore and TRI.

Algorithm 4.

Centrality measures.

Input: G (N, R)
Output: Set T = T_Bc, T_Cc, T_Evc, T_PRc ▹T_Xc = n₁, n₂,…, n_k,

\forall n_{i} \in N

where k = 10, 20, 30
1: Step 1: Calculate B_C(n_i);

\forall n_{i} \in N

2: Step 2: Calculate C_C(n_i);

\forall n_{i} \in N

3: Step 3: Calculate Ev_C(n_i);

\forall n_{i} \in N

4: Step 4: Calculate PR_C(n_i);

\forall n_{i} \in N

N: number of nodes; R: number of relations; T_Bc: top-k list derived from betweenness centrality; T_Cc: top-k list derived from closeness centrality; T_Evc: top-k list derived from eigenvector centrality; T_PRc: top-k list derived from PageRank centrality.

3.5. Collective influence finding

To expansively examine the landscapes of influential users given by each centrality measure, a new list T_com having collective results is generated as displayed in Figure 3.

Figure 3.

Process to find collective influence from centrality-based top-k lists.

Here, T_com is a set of distinct influential nodes n_i which are commonly appeared in T_Bc, T_Cc, T_Evc and T_PRc with occurrence threshold $t_{o} \geq 1$ . The processing steps are displayed in Algorithm 5. Com_Func depicted in line number 1 computes the collective influence list.

Algorithm 5.

Collective influence finding.

Input: Set T = T_Bc, T_Cc, T_Evc, T_PRc
T_Xc = n₁, n₂,…, n_k,

\forall n_{i} \in N

where k = 10, 20, 30
Output: Set T = T_com Where T_com is a set of distinct influential nodes n_i which are present in T_Bc, T_Cc, T_Evc, T_PRc with occurrence threshold

t_{o} \geq 1

1: T_com = Com_Func (T_Bc, T_Cc, T_Evc, T_PRc)
T_com = Distinct(n_i) ▹where

n_{i} \in T_{Bc}, T_{Cc}, T_{Evc}, T_{PRc}

with

t_{o} \geq 1

T_Bc: top-k list derived from betweenness centrality; T_Cc: top-k list derived from closeness centrality; T_Evc: top-k list derived from eigenvector centrality; T_PRc: top-k list derived from PageRank centrality.

The skills, affiliations, departments and locations possessed by the identified influential users in T_com are further derived and analysed in this research. These results are presented in section ‘Experimental analysis and results’.

4. Experimental analysis and results

A summary comprising details of implementation setup is shown in Table 1.

Table 1.

Implementation setup.

Process	Tools and technology
Data collection	Web rendering and XPath
Graph generation	Cypher Query Language
Graph storage	Neo4j
Graph model generation	GraphML
Graph layout generation	Gephi
Centrality measures	BC, CC, EvC, PRC

BC: Betweenness Centrality; CC: Closeness Centrality; EvC: Eigenvector Centrality; PRC: PageRank Centrality.

The rendering process renders the profile content of 1544 RG users. For each user, 18 profile properties are rendered. In total, 27,792 (1544 × 18) properties and 1646 connections present among 1544 users are rendered. Table 2 shows the 18 rendered properties with the assigned IDs.

Table 2.

List of rendered properties.

Property ID	Property attribute	Property ID	Property attribute
P₁	Name	P₁₀	No. of projects
P₂	No. of answers	P₁₁	No. of questions
P₃	No. of citations	P₁₂	RGScore
P₄	Department	P₁₃	No. of reads
P₅	No. of followers	P₁₄	No. of recommendations
P₆	No. of followings	P₁₅	No. of research items
P₇	Location	P₁₆	Total research interest
P₈	Affiliation	P₁₇	No. of skills
P₉	Position	P₁₈	Set of skills {s₁, s₂,..,s_n}

The visual representation of the constructed RG graph is shown in Figure 4. Each RG user is depicted as a node. The rendered 18 properties of each user are stored with each respective node. The rendered follower and following connections are displayed as directed edges with labels ‘Follower’ or ‘Following’ among connected nodes.

Figure 5.

Performance of BC for top-k RG influential users. (a) Performance of BC for top-10 RG influential users.(b) Performance of BC for top-20 RG influential users. (c) Performance of BC for top-30 RG influential users.

Figure 4.

Visualisation of the constructed RG graph.

The performance of standard centrality measures is evaluated based on how efficiently these measures identify top-k influential users from the constructed RG graph. The experimentation is performed taking three values of k with identical intervals,that is, k = 10, 20 and 30. Four centrality measures are implemented on the graph in iterative manner, and eventually, four lists of top-k influential users are recorded. In Table 3, top-k (k = 10) influential RG users received by centrality measures are presented. Here, Rank 1 denotes the highest influence. The commonly found users in all four centrality top-k lists are highlighted in Table 3.

Table 3.

Identified top-k (k = 10) influential users in RG graph.

Rank	BC	CC	EvC	PRC
1	Mirac Yazici	Mirac Yazici	Mirac Yazici	Mirac Yazici
2	Elhanan Helpman	Elhanan Helpman	Elhanan Helpman	Liat Helpman
3	Liat Helpman	H Gin Chong	Murat Aktan	Limor Helpman
4	Limor Helpman	Selvi Yıldız	Gulcin Guresci	Elhanan Helpman
5	H Gin Chong	Devrim Yalcin	Anil Bolukoglu	Murat Aktan
6	Selvi Yıldız	Mohammed Ahmed Endris	Merve Kosaroglu	Gulcin Guresci
7	Devrim Yalcin	Nzeh Emeka C	Meelis Kitsing	Anil Bolukoglu
8	Mohammed Ahmed Endris	Mohamed I Elgohari	Ghaidaa Hetou	Merve Kosaroglu
9	Nzeh Emeka C	Vedat Aktepe	Başak Karşıya-kalı	Meelis Kitsing
10	Mohamed I Elgohari	Murat Aktan	Üzeyir Aydın	Ghaidaa Hetou

BC: betweenness centrality; CC: closeness centrality; EvC: eigenvector centrality; PRC: PageRank centrality.

To identify the relevance of RG followers and followings with respect to RGScore and TRI, top-k lists generated by centrality measures are compared with the top-k lists computed by RGScore and TRI.

Table 4 shows the sample ranking tuples generated for every centrality with RGScore and TRI for top-k results. The generated ranking tuples disclose the degree of relevance between two ranking lists. $< {BC}_{Rki}, {RG}_{Rki} >$ denotes the ranking tuple for BC with RG for rank i and $< {BC}_{Rki}, {TRI}_{Rki} >$ denotes the ranking tuple for BC with TRI for rank i. Here, i = 10, 20 and 30 for top-10, top-20 and top-30, respectively. For each centrality measure, such tuples are generated.

Table 4.

Ranking tuples notations for top-k.

Centrality	RGScore tuple	TRI tuple
BC	$< {BC}_{Rki}, {RG}_{Rki} >$	$< {BC}_{Rki}, {TRI}_{Rki} >$
CC	$< {CC}_{Rki}, {RG}_{Rki} >$	$< {CC}_{Rki}, {TRI}_{Rki} >$
EvC	$< {EvC}_{Rki}, {RG}_{Rki} >$	$< {EvC}_{Rki}, {TRI}_{Rki} >$
PRC	$< {PRC}_{Rki}, {RG}_{Rki} >$	$< {PRC}_{Rki}, {TRI}_{Rki} >$

RG: ResearchGate; TRI: Total Research Interest; BC: betweenness centrality; CC: closeness centrality; EvC: eigenvector centrality; PRC: PageRank centrality.

The generated ranking tuples for $< BC, RGScore >$ and $< BC, TRI >$ for top-10, top-20 and top-30 influential RG users are displayed in Figure 5(a)–(c), respectively. Similarly, ranking tuples for $< CC, RGScore >$ and $< CC, TRI >$ , for top-10, top-20 and top-30 influential RG users are displayed in Figure 6(a)–(c), respectively. For $< EvC, RGScore >$ and $< EvC, TRI >$ , top-10, top-20 and top-30 influential RG users’ ranking tuples are shown in Figure 7(a)–(c), respectively. In Figure 8(a)–(c), ranking tuples for $< PRC, RGScore >$ and $< PRC, TRI >$ for top-10, top-20 and top-30 influential RG users are displayed, respectively.

Figure 6.

Performance of CC for top-k RG influential users. (a) Performance of CC for top-10 RG influential users. (b) Performance of CC for top-20 RG influential users. (c) Performance of CC for top-30 RG influential users.

Figure 7.

Performance of EvC for top-k RG influential users. (a) Performance of EvC for top-10 RG influential users. (b) Performance of EvC for top-20 RG influential users. (c) Performance of EvC for top-30 RG influential users.

Figure 8.

Performance of PRC for top-k RG influential users. (a) Performance of PRC for top-10 RG influential users. (b) Performance of PRC for top-20 RG influential users. (c) Performance of PRC for top-30 RG influential users.

To identify the correlations among the generated ranking tuples, rank correlation co-efficient is computed. The results are demonstrated in Figure 9.

Figure 9.

Ranking correlation measures. (a) Ranking correlation for top-10. (b) Ranking correlation for top-20. (c) Ranking correlation for top-30.

It is evident from Figure 9 that for both RGScore and TRI, BC performs efficiently while EvC and PRC deliver inefficient results for the rendered data and constructed RG graph. Higher correlation in the case of BC suggests that high number of RG users depends on one another to make connections. Wider set of RG users in the constructed graph has many connections that are very crucial in the propagation. EvC and PRC achieve lower correlation which depicts overall less number of RG users in each user’s research network. It is observed that for k = 10 and k = 20, BC outperforms, while for k = 30, CC achieves higher correlation when compared with both RGScore and TRI. Overall, CC performs efficiently over EvC and PRC, which suggests in the constructed graph, a moderate set of RG users has ties to several active or influential users. Overall, as the k in top-k increases, the performance of EvC and PRC decreases. BC and CC perform stable with increasing k values, demonstrating BC as the most relevant to both RGScore and TRI for the constructed RG graph in influence identification and ranking yielding the average relevance of 67% with RGScore and 40% with TRI.

4.1. Collective influence list

Based on the collective influence finding method presented in Algorithm 5, for the influential users present in each top-k list of four centrality measures, the users with occurrence threshold of at least 1 are fetched. The properties of such derived 59 users are examined in detail to identify their skills, degree affiliations, departments and locations.

Figure 10(a) represents the influential skills of users present in T_com list and having occurrence count $C_{o} \geq 5$ . The skill IDs to skills mapping are S₁: Macroeconomics, S₂: Development Economics, S₃: International Economics, S₄: Financial Economics, S₅: Econometrics, S₆: Microeconomics, S₇: Monetary Economics, S₈: Qualitative Social Research, S₉: Quantitative Social Research, S₁₀: Social Policy, S₁₁: Business Administration, S₁₂: Economics, S₁₃: Political Economy, S₁₄: Public Administration, S₁₅: Economic History, S₁₆: Environmental Economics, S₁₇: Public Economics. Similarly, influential degree affiliations, departments and locations that are present in T_com list and having occurrence count $Co > 1$ are depicted in Figure 10(b) to (d), respectively.

Figure 10.

Property analysis of top-k RG users in T_com list. (a) Skills. (b) Degree affiliation. (c) Department. (d) Location.

4.1.1. Findings

The skills of identified common influential RG users are examined in-depth to gain knowledge about in which research fields the influential researchers are working. It is recognised that areas, such as Macroeconomics, Development Economics, International Economics, Financial Economics and Econometrics are highly significant with occurrence count $Co \geq 10$ in the T_com list. Microeconomics, Monetary Economics, Qualitative Social Research, Quantitative Social Research, Social Policy, Business Administration, Economics, Political Economy, Public Administration, Economic History, Environmental Economics and Public Economics have average significance with occurrence count $Co \geq 5$ and occurrence count $Co < 10$ . Behavioural Economics, Economic Growth, Financial Risk Management, Institutional Economics, International Political Economy, Banking and Finance and Business Economics are identified as less significant with occurrence count $Co \geq 1$ and occurrence count $Co < 5$ in the T_com list, respectively.

In the computed T_com list, majority influential users possess PhD in ‘Economics’ with $Co \geq 20$ . The others pursuing PhD in ‘Economics’ or MSc in ‘Economics’ with equal $Co$ of 2. Other affiliated degrees of influential users are revealed to be MA, MBA, MPhil and MSS with $Co = 1$ .

The influential university departments found in T_com are Department of Economics with $Co = 17$ ; Department of Accounting and Finance, Department of International Trade and Finance, Department of Research with $Co = 2$ . Other departments such as Department of Accounting, Finance and MIS, Department of Agricultural Economics, Department of Banking, Department of Economics and Finance, Department of Environmental Engineering, Department of International Economics, Department of Land Economy, Department of Obstetrics and Gynaecology, Department of Political Science and Economics and Finance Department with $Co = 1$ were present in T_com.

The observed influential locations in T_com list are İzmir and Nevşehir with $Co = 7$ ; Istanbul with $Co = 4$ ; Ankara and Muğla with $Co = 2$ . In common, these locations belong to Turkey. Other identified less significant locations in T_com list are Addis Ababa, Ethiopia; Al Manşūrah, Egypt; Arba Minch, Ethiopia; Artvin, Turkey; Berlin, Germany; Brisbane, Australia; Pune, India; Washington, DC, United States; Enugu, Nigeria; Niğde, Turkey; Desē, Ethiopia; Bursa, Turkey; Cambridge, United Kingdom; Cambridge, United States; Genève, Switzerland; Katra, India; Kfar Saba, Israel; Leeds, United Kingdom; Melikgazi, Turkey; Mīzan Teferī, Ethiopia; New Brunswick, United States; Prairie View, United States; Rouen, France; Sivas, Turkey; Tallinn, Estonia; Tel Aviv, Israel; Trondheim, Norway and Xi’an, China with $Co = 1$ . Overall, Turkey is found to be a location where majority of influential RG users with significant research in ‘Economics’ belong as per the rendered data.

5. Summary and future work

The proposed methodology provides a novel approach to treat RG information as a connected graph of users in terms of followers and followings. The aim is to anatomise the impact of RG followers and followings on influence identification. As the already established graph of RG data is not available, research demands an efficient method to collect linked RG information to form RG graph. The methodology contributes to offer a data rendering process to fetch linked RG information. From the fetched information, RG graph is constructed. Standard centrality measures – betweenness, closeness, eigenvector and PageRank – are implemented on the constructed graph for influence identification. The performance is compared with RGScore and TRI. The results reveal that BC and CC perform efficiently on our constructed RG graph. This research is an initial attempt to analyse RG information as a complete and connected graph by implementing centrality measures with an aim to study the impact of follower and following connections on influence identification.

ASNS data undergo less frequent changes unlike social media data; hence, this research provides a static pattern that does not consider the change in centrality measures over a time at this stage. We aim to consider the effects of the topological changes on influence in our future work. The dynamic changing patterns in the graph will be helpful to study symmetry of the complete RG graph and grasp deeper insights into influence change over a time. In addition, the centrality measures applied on RG graph do not consider the statistical measures of RG such as number of research items, citations, reads, etc. In the future, centrality measures and RG statistics can be assembled to identify the influence more accurately.

Footnotes

Appendix 1

Table 6.

List of notations.

U	ResearchGate User
N	Number of Nodes
N_i	ith node
P	Number of properties
P_i	ith property of node N
R	Number of relations
FL	Follower of node N
FW	Following of node N
I_scn	Influence score of node n, where $n \in N$
TB_c	Top-k list derived from betweenness centrality
TC_c	Top-k list derived from closeness centrality
TEv_c	Top-k list derived from eigenvector centrality
TPR_c	Top-k list derived from PageRank centrality
T_com	Collective top-k list derived from centrality results
t_o	Occurrence threshold
C_o	Occurrence count
S_i	ith skill of node N

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Mitali Desai

References

Zhang

. Finding high-impact interdisciplinary users based on friend discipline distribution in academic social networking sites. Scientometrics 2019; 119(2): 1017–1035.

Ansari

JAN

Khan

. Exploring the role of social media in collaborative learning the new domain of learning. Smart Learn Environ 2020; 7(1): 1–16.

Mas-Bleda

Thelwall

Kousha

et al. Do highly cited researchers successfully use the social web? Scientometrics 2014; 101(1): 337–356.

Yan

Zhang

. Research universities on the ResearchGate social networking site: an examination of institutional differences, research activity level, and social networks formed. J Inform 2017; 12(1): 385–400.

Hoffmann

Lutz

Meckel

. A relational altmetric? Network centrality on ResearchGate as an indicator of scientific impact. J Assoc Inform Sci Technol 2015; 67(4): 765–775.

Kelly

. Using social media to enhance your research activities. In: Proceedings of the social media in social research 2013 conference, https://purehost.bath.ac.uk/ws/portalfiles/portal/11393548/sra_2013.pdf

Amjad

Daud

Aljohani

. Ranking authors in academic social networks: a survey. Libr Hi Tech 2018; 36(1): 97–128.

Fan

Yuan

. Research on pathways of expert finding on academic social networking sites. Inform Process Manage 2021; 58(2): 102475.

Liang

Mao

et al. Combining deep neural network and bibliometric indicator for emerging research topic prediction. Inform Process Manage 2021; 58(5): 102611.

10.

Nerur

Sikora

Mangalaraj

et al. Assessing the relative influence of journals in a citation network. Commun ACM 2005; 48(11): 71–74.

11.

Iyengar

Balijepally

. Ranking journals using the dominance hierarchy procedure: an illustration with IS journals. Scientometrics 2015; 102(1): 5–23.

12.

Zhang

Zhao

Cheng

et al. Finding influential papers in citation networks. In: Proceedings of the IEEE 1st international conference on data science in cyberspace (DSC), Changsha, China, 13–16 June2016, pp. 658–662. New York: IEEE.

13.

Bai

Zhang

Hou

et al. Quantifying the impact of scholarly papers based on higher-order weighted citations. PLoS ONE 2018; 13(3): e0193192.

14.

Savov

Jatowt

Nielek

. Identifying breakthrough scientific papers. Inform Process Manage 2020; 57(2): 102168.

15.

Ortega

. Toward a homogenization of academic social sites: a longitudinal study of profiles in Academia.edu, Google Scholar Citations and ResearchGate. Online Inform Rev 2017; 41: 812–825.

16.

Jordan

. From social networks to publishing platforms: a review of the history and scholarship of academic social network sites. Front Digit Human 2019; 6: 5.

17.

Rapple

. Understanding and supporting researchers’ choices in sharing their publications: the launch of the fairshare network and shareable pdf. Insights 2018; 31: 408.

18.

Gasparyan

Nurmashev

Yessirkepov

et al. Researcher and author profiles: opportunities, advantages, and limitations. J Korean Med Sci 2017; 32(11): 1749–1756.

19.

Hashmi

HBUH

Kayani

HUR

Toor

et al. The impact of social media: a survey. Int J Sci Technol Res 2020; 9(12): 341–348.

20.

Bamakan

SMH

Nurgaliev

Opinion leader detection: a methodological review. Exp Syst Appl 2018; 115: 200–222.

21.

Silva

LOJE

Maldonado

Brigham

et al. Evaluating scholars’ impact and influence: cross-sectional study of the correlation between a novel social media–based score and an author-level citation metric. J Med Intern Res 2021; 23(5): e28859.

22.

Desai

Mehta

Rana

. An empirical analysis to identify the effect of indexing on influence detection using graph databases. Int J Innov Technol Explor Eng 2019; 8(9S): 414–421.

23.

Bromham

Dinnage

Hua

. Interdisciplinary research has consistently lower funding success. Nature 2016; 534(7609): 684–687.

24.

Elsayed

. The use of academic social networks among Arab researchers: a survey. Soc Sci Comput Rev 2016; 34(3): 378–391.

25.

Ding

. Scientific collaboration and endorsement: network analysis of coauthorship and citation networks. J Inform 2011; 5(1): 187–203.

26.

Mangan

. Social networks for academics proliferate, despite some doubts. Chron High Educ 2012; 58(35): 1–7.

27.

Desai

Mehta

Rana

. Issues and challenges in big graph modelling for smart city: an extensive survey. Int J Comput Intell IoT 2019; 1(1): 2018.

28.

Petkova

. Why network databases make a better home for interconnected data than the relational databases, 2016, https://www.ontotext.com/blog/graph-databases-interconected-data-relational-databases/ (accessed 11 June 2018).

29.

Patel

Dharwa

. Graph data: the next frontier in big data modeling for various domains. Indian J Sci Technol 2017; 10: 1–7.

30.

Srinivasa

. Data, storage and index models for graph databases. Hershey, PA: IGI Global, 2011, pp. 4770–4796.

31.

Malinen

Koivula

. Influencers and targets on social media: investigating the impact of network homogeneity and group identification on online influence. First Monday 2020; 25: 10453.

32.

Valsesia

Proserpio

Nunes

. The positive effect of not following others on social media. J Market Res 2020; 57(6): 1152–1168.

33.

Landherr

Friedl

Heidemann

. A critical review of centrality measures in social networks. Bus Inform Syst Eng 2010; 2: 371–385.

34.

Iacobucci

McBride

Popovich

et al. In social network analysis, which centrality index should I use? Theoretical differences and empirical similarities among top centralities. SSRN Electron J 2017; 8: 72–99.

35.

Zhang

Luo

. Degree centrality, betweenness centrality, and closeness centrality in social network. In: Proceedings of the 2nd international conference on modelling, simulation and applied mathematics (MSAM2017), https://doi.org/10.2991/msam-17.2017.68

36.

Bródka

Musial

Kazienko

. A performance of centrality calculation in social networks. In: Proceedings of the 2009 international conference on computational aspects of social networks, Fontainebleau, 24–27 June2009, pp. 24–31. New York: IEEE.

37.

Wan

Mahajan

Kang

et al. A survey on centrality metrics and their network resilience analysis. IEEE Access 2021; 9: 104773–104819.

38.

Matas

Martincić-Ipšić

Meštrović

. Comparing network centrality measures as tools for identifying key concepts in complex networks: a case of Wikipedia. J Digit Inform Manage 2017; 15(4): 203–213.

39.

Boudin

. A comparison of centrality measures for graph-based keyphrase extraction. In: Proceedings of the 6th international joint conference on natural language processing, https://hal.archives-ouvertes.fr/hal-00850187/document

40.

Hua

Huang

. Centrality metrics’ performance comparisons on stock market datasets. Symmetry 2019; 11: 916.

41.

Yan

Ding

. Applying centrality measures to impact analysis: a coauthorship network analysis. J Am Soc Inform Sci Technol 2010; 60: 1–17.

42.

Cardente

. Using centrality measures to identify key members of an innovation collaboration network, Technical Report, 2012, http://snap.stanford.edu/class/cs224w-2012/projects/cs224w-043-final.pdf

43.

Amjad

Daud

Akram

et al. Impact of mutual influence while ranking authors in a co-authorship network. Kuwait J Sci 2016; 43(2): 11–19.

44.

Amjad

Daud

Che

et al. MuICE: mutual influence and citation exclusivity author rank. Inform Process Manage 2016; 52(3): 374–386.

45.

Bibi

Khan

Iqbal

et al. Ranking authors in an academic network using social network measures. Appl Sci 2018; 8(10): 1824.

46.

Ghane

Zowj

Ehsanifar

. Identifying information retrieval research trends using author co-citation network. Int J Inform Sci Manage 2019; 17: 99–117.

47.

Maia

LFMP

Lenzi

Rabello

et al. Scientific collaboration in Zika: identification of the leading research groups and researchers via social network analysis. Cad Saude Publica 2019; 35: e00220217.

48.

Makkizadeh

Dehghan

Mostafavi

. Investigating association between social influence, productivity, and performance in co-author network of researchers in medical ethics. Iran J Med Ethic Hist Med 2020; 13: 240–252.

49.

Wang

Ding

Wei

et al. Scholars influence evaluation based on time series heterogeneous network. In: Proceedings of the 13th international conference on measuring technology and mechatronics automation (ICMTMA), Beihai, China, 16–17 January2021, pp. 868–871. New York: IEEE.

50.

Kesht-karan

Ghane

Danesh

. Dimensions of the scientific collaborations of the researchers affiliated with Shiraz University of Medical Sciences. Int J Inform Sci Manage 2021; 19(2): 31–48.

51.

Yaghtin

Sotudeh

Mohammadi

et al. A correlation study of co-opinion and co-citation similarity measures. Int J Inform Sci Manage 2019; 17(2): 19–31.

52.

Samie

Biranvand

Rahmaniyan

et al. The impact of the activity of industrial engineering researchers in various scientific-citation networks on improving their scientific authority status. Int J Inform Sci Manage 2022; 20(1): 257–271.

53.

Desai

Mehta

Rana

. RGNet: the novel framework to model linked ResearchGate information into network using hierarchical data rendering. In: Patnaik

Yang

Sethi

(eds) Advances in machine learning and computational intelligence. Singapore: Springer, 2021, pp. 37–45.

54.

Nicholas

Clark

Herman

. ResearchGate: reputation uncovered. Learn Publ 2016; 29(3): 173–182.

55.

Keller

. Information and tips related to search engines like Google Scholar and other ways your work is and can be more visible on the net, 2019, https://www.researchgate.net/publication/332593431_Information_and_tips_related_to_search_engines_like_Google_Scholar_and_other_ways_your_work_is_and_can_be_more_visible_on_the_net

56.

Boudry

Durand-Barthez

. Use of author identifier services (ORCID, ResearcherID) and academic social networks (Academia.edu, ResearchGate) by the researchers of the University of Caen Normandy (France): a case study. PLoS ONE 2020; 15(9): e0238583.

57.

Hammook

Mišić

Misic

. Crawling ResearchGate.net to measure student/supervisor collaboration. In: Proceedings of the IEEE global communications conference (GLOBECOM), San Diego, CA, 6–10 December2015.

58.

Thelwall

Kousha

. ResearchGate: disseminating, communicating, and measuring scholarship? J Am Soc Inform Sci Technol 2014; 2014: 876–889.

59.

Kadriu

. Discovering value in academic social networks: a case study in ResearchGate. In: Proceedings of the ITI 35th international conference on information technology interfaces, Cavtat, 24–27 June2013, pp. 57–62. New York: IEEE.

60.

TZJ

Song

Chiu

DM.

The academic social network. Scientometrics 2014; 101(1): 203–239.

61.

Khvatova

Dushina

. Scientific online communication: the strategic landscape of ResearchGate users, 2021, https://www.researchgate.net/publication/350239194_Scientific_Online_Communication_The_Strategic_Landscape_of_ResearchGate_Users

62.

Khvatova

Dushina

Nikolaenko

. Do the online activities of scientists in social professional networks influence their academic achievements? In: Proceedings of the European conference on management, leadership and governance, https://www.researchgate.net/publication/321768233_Do_the_Online_Activities_of_Scientists_in_Social_Professional_Networks_Influence_their_Academic_Achievements

63.

Heydari

Teimourpour

. Analysis of ResearchGate, a community detection approach. In: Proceedings of the 6th international conference on web research (ICWR), Tehran, Iran, 22–23 April2020, pp. 319–324. New York: IEEE.

64.

Yan

Zhang

et al. How does scholarly use of academic social networking sites differ by academic discipline? A case study using ResearchGate. Inform Process Manage 2021; 58(1): 102430.

65.

Niu

Huang

. Information behavior of online academic social platform: case study of ESI highly cited researchers of ResearchGate. In: Proceedings of the 4th international conference on software engineering and information management, Yokohama, Japan, 16–18 January2021, pp. 158–164. New York: ACM.