The complexity of this task comes in the form of creating this dataset with the use of the semantic web and web ontology languages. And also in drawing conclusions and creating data visualisation of the network. For this I opted for Gephi for it's ability to compute various graph metrics, and a Geographic Information Systems (GIS) application to draw on international relations in co-authorship.
Without prevalent web ontology frameworks being implemented for public deployement; to a certain extent, it is possible to web scraping with website specific knowledge to gather publically available information to then use several API's to develop a corpus of publication metadata which can then be used for any number of analyses.
The specific methodology employed was:
- Gather the names of researchers listed as being members of a department at a given university
- Gather their ORCID Identifiers where publicly listed (e.g. staff profile pages).
- Use their ORCID identifiers or name (if for any reason ORCID identifiers are unknown), to gather authorship attribution identifier for this researcher from SCOPUS.
- Use SCOPUS's attribution identifier to enumerate all the works of a researcher.
- Infer the SCOPUS authorship identifiers of all their co-authors.
- Infer all the co-authors' ORCID Identifiers.
- Gather individual co-author metadata through all available means (ORCID and SCOPUS), chiefly the most recent institution, the country of this institution and researcher's fields of interest.
- Dataset is prepared
I then use the data gathered to generate a graph
The datasets cannot be made public to respect the privacy of the real researchers about whom this data was gathered. and unfortunately annonymisation is also not an option as this is also not allowed under the terms of the Scopus and ORCID API's.
Analysing trends in the co-authorship activity of Durham Computer science researchers using collaborative social networks
The chosen modelling task is to examine co-authorship in the department of computer Science at Durham university. The chosen implementation was a social network which consists of nodes which represent researchers, and a single edge between two nodes represents a co-authorship collaboration on a single publication. Social Network analysis (SNA) is the use of networks within the context of graph theory, to model some real-world behaviour; by modelling co-authorship as a social network, it is possible to correlate the structure of the network to some constraint in the environment. SNA's key applications are in the Humanities and social sciences from examining culture spreading to information circulation. SNA has proven to provide meaningful insight in evaluating research networks [1]. It has also been shown in studies that use bibliometric data that co-authored papers achieve more exposure and impact [2],
Since there is no publicly available dataset one had to be created first; the data was gathered in 3 steps. The first step involved scraping the names and ORCID details of each researcher in the department using the Durham university website; most researchers from the Computer Science department's website. ORCID is a 16-digit number in the format xxxx-xxxx-xxxx-xxxx, which uniquely identifies a certain researcher but requires them to create an account first.
This was an issue since many did not; additionally, ORCID only displays publications the researcher wishes to make public; in order to gain access to the corpus of all of their works a different database would be required. SCOPUS is one such corpus with unique author attribution for all of the researcher's works that are publicly available. Scopus uses a rich underlying metadata architecture to index its corpus of publications and their authorship information to make searching unstructured meta data feasible [3].
The SCOPUS API has several layers of security, access can only be made through institutional IP addresses, and an API key must be registered with a cap of 7,500 queries. The API contains all the data required however was not built with this specific use case; and so the process of obtaining the dataset used was quite convoluted. The dataset comprises of 2,400 researchers collaborating on 2,300 publications, forming a co-authorship social network.
Authorship data can be found using ORCID numbers or even just their name and a Durham university affiliation. All scopus author IDs were gathered for our staff members, in step 2 The SCOPUS database was queried with these IDs, and Scopus's authorship attribution fetched all of the researcher's publications, along with the co-author information. Step 3 is gathering affiliation data for each co-author, their names, university affiliation and country where they are currently active.
The two csv files in this project comprises the built dataset. Articles.csv consists of pairwise undirected edges of two researcher's unique IDs which represent collaboration on a single publication, if they collaborate in multiple papers, they will have more edges. Researchers.csv consists of researcher's affiliation information as well as their names to validate the uniqueness of the author IDs taken from scopus.
The nodes and edges were imported into Gephi as two csv files and were formatted using the Force Atlas and Yifan Hu layouts, the toolkit also produced all the metrics seen in this paper. The visualisations shown in this paper include the manipulation of feature data for each node, in order to produce a word cloud that gives context to the clusters. The names in these visualisations are not anonymised since this data is publicly available, however some ethical considerations should have been made to anonymise the data collected.
Figure 1.0 The co-authorship social network of Durham researchers with red edges showing collaboration within the department.
Examining (Figure 1.0), the data suggests the specific areas of interest that have little intra-department collaboration, are Virtual/Augmented reality, graph theory, NLP, high performance computing and cryptography. This is done by observing the clustering of certain modularity classes, and the Durham researchers within the clique and confirming they encompass the whole department's contribution to this topic, it is also clear these modularity classes are weakly connected components, and such it is possible to draw this conclusion from (Figure 1.0). It is typical in social networks for the tendency to form communities through triadic closure, however it is seen that weakly connected and disjoint components reflect the nicheness of the researcher's topic of interest in not collaborating often within the department.
Figure 2.0 The co-authorship social network coloured by modularity class with labelled for Durham researchers and scaled in size by the author's publication count.
The network depicted above was coloured using the modularity classes created by using Blondel's algorithm to depict the different clusters in the graph. The size of each label shows the degree centrality of that node. Modularity is the deviation from the expected random proportion of edges within communities, and the assignment is made iteratively converging on some reasonable objective function.
Intuitively the model appears to show that some staff members particularly recent additions have their own neighbourhood of researchers with whom they form a clique which other researchers in the department do not interact with. And a main cluster of long-standing staff members who collaborate more often and collaborate with each other's co-collaborators. It is also clear that there is a sizeable cluster of PhD students who co-author a paper with certain professors and rarely contribute to any works outside of the university.
Certain researchers do not have many publications with internal co-authoring, this is most likely due to the length of their career and being less research oriented within the department as it stands with David Budgen and software engineering being a specific example from the graph above. One limitation of the data gathered toward this model is that it does not consider the publications made only in the researcher's tenure at Durham, and so the graph is populated with historical connections that are likely not ongoing co-authorships, and this is particularly evident from the clusters in each corner of (Figure 1.2 ).
Fields such as computer vision and specific AI applications lead to some clear clustering around certain researchers, this is likely due to co-authorship in PhD research, one such example is the blue clustering in the centre of the Figure X-1.0. One other observation is the green cluster directly above.
Figure 1.2 The co-authorship social network coloured by country where the author was last affiliated with, and labels for Durham researchers which are scaled in size by the author's publication count.
When trying to understand a pattern in international collaboration, the model does not present a coherent message. This is as a result of insufficient data mining which is down to how the Scopus API was created, an author does not have to explicitly create a profile for one to exist; and so, for some co-authors where they have not declared a country of activity the default location is the country where their last publication was made publicly available. It is typically either the United States or Switzerland.
Additionally, a researcher's location may have changed since the publication was made and the affiliation attribution was set, which further invalidates any conclusions without further verifying the integrity of the dataset. It is possible however to report the demographic data as collected, with the assumption that researchers typically remain working in the same country.
One missed opportunity for this implementation is the potential use of TEI data, this is because scopus only made university's addresses available on the API the week prior to this assignment's due date; however, had this been available it would have been possible to view international collaborations without using a forced based layout seen in this paper's visualisations. TEI gives a much more intuitive visualisation of node's specific interactions with the rest of the world.
The heatmap shown in (Figure 3.0) shows a large majority of Europe, the Americas and most of Asia being countries whose researchers have collaborated with Durham researchers, but also shows few are disproportionately higher than others. This may be down to the nature of research in the developed world, and the clustering of technology researchers in the parts highlighted. However the conclusion drawn from this model is that since other countries could have a higher degree of centrality in a global co-author social network and display more world wide collaboration; that Durham Computer Science represents a small footprint in such a network.
Table 1.0 showing the demographics of co-authors in this social network.
Figure 3.0 A heatmap of co-author citations with Durham publications.
Figure X-2.6 A further breakdown of the heatmap in Figure X-2.5.
Figure 4.0 The co-authorship social network where the size of a node is determined by it's betweenness centrality
Betweenness centrality measures how central individuals are within the
social network by measuring the path lengths that would flow through this one node compared to all others. This can be seen for the network in (Figure 4.0). There are 2,400 nodes with 2,300 Articles that form 44,850 edges in the network. In calculating the betweenness centrality for this network we find the average path length is of 6. The nodes with highest centrality can be explained, researchers in certain disciplines such as Image processing, are likely to be co-authoring papers with game developers, computer vision and AI researchers; it is evident such disciplines exist when you view these node's topics of interest. And by inspecting the individuals with high centrality we examine their role in facilitating conversation. Nodes with high amount of centrality play a large role in facilitating communication in this case across key topic interests [4].
Degree centrality is a ranking of nodes by their degree. The topology of the graph leads to a cycle starting with one author and passes through each co-author representing a single publication. And so, the degree centrality only highlighted certain Durham researchers who have had long careers and provided no further insight into group dynamics.
{width="2.86587489063867in"
height="1.9069772528433946in"}
{width="2.935416666666667in"
height="1.953251312335958in"}
{width="2.86587489063867in"
height="1.9069772528433946in"}
{width="2.9357731846019246in"
height="1.9534886264216973in"}
{width="2.918297244094488in"
height="1.9418602362204724in"}
Figure 5.0 The graphs produced by blondel's and Louvain algorithms [5] [6] in calculating Betweenness centrality and mean clustering coefficient.
The clustering coefficient explains the triangulation within a network by averaging the clustering coefficients of all its nodes. It is used in the analysis of social networks to measure the degree to which nodes in a graph tend to cluster together [7]. The clustering coefficient of 0.881 suggests the department's researchers have a high tendency to cluster together on related tasks. The local cluster efficient (Watz, Strogaatz) [8] is a measure of to what degree is a node's neighbourhood fully connected, and the global value in an undirected graph such as this is the average local coefficient across all nodes. The number of communities that maximises modularity is 35, with a score of 0.67. And as discussed in the lectures it is apparent that such a large number of clusters despite maximising modularity should be constrained in order to identify fewer key topics in computer science, in order to gather some more data.
One unique observation is the Average path length of 5.9 coinciding with the small world phenomenon [9] of every person's 6 degrees of separation. This would suggest this social network is no more well connected than one of the world's computer science researchers. The small world phenomenon is typically a small average path length and a large global clustering coefficient.
One final limitation of the model is the assumption that all collaborations can be shown through widely available publications such as articles; and fails to capture works such as books or software, this could be strengthened by including more data sources.
Bibliography
| [1] | V. S. Malbas, "Mapping the collaboration networks of biomedical research in Southeast Asia," 2015. |
|---|---|
| [2] | G. &. D. C. Abramo, "(2015). The relationship between the number of authors of a publication...," Journal of Informetrics., 2015. |
| [3] | [Online]. Available: https://www.elsevier.com/solutions/scopus/how-scopus-works#. |
| [4] | M. G. |
| [5] | J.-C. D. M. B. R. Lambiotte, "Laplacian Dynamics and Multiscale Modular Structure in Networks," 2009. |
| [6] | J.-L. G. R. L. E. L. Vincent D Blondel, " Fast unfolding of communities in large networks,," in in Journal of Statistical Mechanics: Theory and Experiment , 2008. |
| [7] | M. E. J. Newman, "Networks: An Introduction," in Oxford University Press, New York, 2010. |
| [8] | D. J. W. &. S. H. Strogatz, "Collective dynamics of 'small-world' networks," Department of Theoretical and Applied Mechanics, Kimball Hall, Cornell University, Ithaca, New York 14853, USA, 1998. |
| [9] | S. Milgram, "The Small World Problem," in Psychology Today. Ziff-Davis Publishing Company, 1967. |