Network construction: We use the degree of overlap in reference lists to construct a network of all articles in our
corpus (a technique called bibliographic coupling). If two articles have non-overlapping reference lists, they are not
linked; if they have identical references, the strength of their connexion is maximal. The network is dynamic because a
moving time window with a length of five years slides through time, one year at a time, starting with window 1950-1954,
then moving to 1951-1955 and so on until the end of our period in 2012-2016. This definition of our network
means that two documents that have been published more than five years apart are in the same overall network, but
never appear jointly in a time window.
Cluster detection: In each time window, an algorithm for community detection based on
modularity optimization is applied to the current state of the network. The algorithm used is a variation on the
Louvain method. The panel depicting linked circles is a
representation at the level of these communities (not of the underlying documents). Once we have community
assignments for each window, basic rules based on proximity of clusters from one time window to the next are applied
to determine whether a cluster survives, splits, merges or dies without leaving a trace. These rules generate the polygons
(i.e. the specialties) that stretch from the beginning to the end of our period in the first panel.
Keywords retrieval: Keywords represent the words that most characteristically distinguish the titles of the articles in
the relevant cluster from articles in the other clusters. The panel with polygons gives keywords for each cluster over its
full lifetime. The panel with linked circles gives time-specific keywords. These keywords slowly change through time as
a specialty changes its focus.
Want more information on the procedure?
The article for our project on the social sciences and humanities is in progress at the moment, but our procedure is highly similar
to what we did in our project on economics.
See this article
(and its Appendix for technical details).
Team and Acknowledgments
This project was made possible by a SSHRC Insight Development Grant [430-2016-00758].
was project leader, with Yves Gingras as cocandidate.
The main research assistants on the project are
(from the BIN),
and Philippe Boissonneault (from the
en humanités numériques, Université de Sherbrooke).