Fig 1.
Scopus’ intermediate classification.
List of intermediate classifications for journal labeling according to Scopus, grouped by macro area.
Fig 2.
Top left: yearly number of new publication entries (a publication is counted multiple times if it is listed under more than one profile). This metric has been increasing since the mid-90s; decrease in last couple of years may be due to right-censoring. Top right: yearly number of publications entries normalized by number of authors. Between 1990 and 2016, this number has increased nearly 50%. Bottom: Complementary Cumulative Distribution Function (CCDF) of the number of publications per scientist and institution. Both distribution are heavy tailed, but the distribution for institutions seems to follow a power law with cutoff at x = 105.
Table 1.
For each macro area X shown in the rows, columns report the number of publications entries on venues associated with X, number of researchers who authored one or more publications on such venues, the average number of papers published by these authors on those venues and the average number of sub-areas associated with those publication entries.
Fig 3.
These networks are created using the publication records in [2000, 2014]. Each node is a research field, color-coded according to the respective macro area. We first include the edges of maximum spanning tree and then add all edges whose weights are above a threshold p arbitrarily chosen for better visualization. Left: Frequentist model, p = 0.212. Right: Embedding model, p = 0.35.
Fig 4.
Fraction of subfields in each of the intermediate fields in which some of the institutions are specialized (data from the time interval [2000, 2014]). Intermediate classifications are grouped and colorcoded according to the macro field. The Multidisciplinary (MULT) intermediate field has only a single subfield and is then shown with hatches.
Fig 5.
Distribution of number of active and developed fields.
Complementary Cumulative Distribution Function of the number of active/developed fields for scientists and institutions (data from years [2000, 2014]).
Fig 6.
AUROC’s violin plots for all transitions.
Violin plots of the AUROC for the predicted transitions for scientists, institutions and states. For each plot, the black horizontal line shows the median and the white circle shows the mean. Time windows are set as: model fitting, [1999, 2013], density estimation, [2011, 2013], and prediction testing, [2014, 2016].
Table 2.
Average AUROC for each transition, first setup.
Average AUROC of the predicted transitions for scientists, institutions and states. Time windows are set as: model fitting, [1999, 2013], density estimation, [2011, 2013], and prediction testing, [2014, 2016]. Entities that did not transition during the [2014, 2016] were not included in the sample.
Fig 7.
AUROC’s boxplots for scientists’ transitions (various periods).
Boxplots of the AUROC for the 0 → A transition for scientists. The orange horizontal line shows the median and the white circles shows the mean. Time windows for both model fitting and density estimation are shown in the x axis (Δϕ = ΔRCA = 2). Time windows for prediction testing begin immediately after the previous ones (Δtest = 2).
Table 3.
Average AUROC for scientists 0 → A transition, second setup.
Average AUROC of the predicted 0 → A transition for scientists. Time windows for both model fitting and density estimation are shown in the column headers (Δϕ = ΔRCA = 2). Time windows for prediction testing begin immediately after the previous ones (Δtest = 2). Entities that did not transition during the time window were not included in the sample.
Fig 8.
AUROC distribution for institutions vs. data properties.
Empirical distributions of the AUROC as a function of the logarithms of the (i) number of active fields, (ii) number of developed fields and (iii) normalized number of papers. Time windows are set as in the first experimental setup. Contour plots were created from 500 sample points using Gaussian kernels.
Fig 9.
AUROC’s Coeff. of variation for institutions vs. data properties.
Ratio between sample standard deviation and sample mean of the AUROC as a function of the (i) number of active fields, (ii) number of developed fields and (iii) normalized number of papers. Time windows are set as in the first experimental setup. Statistics are computed using a sliding window of size 1000 on the x axis.
Fig 10.
First, second and third quartiles of AUROC yielded by the Frequentist model for different prediction testing windows, when predicting the transition inactive to active for the 15,491 scientists that performed some transition in every window. Time windows for model fitting and density estimation are resp. [1990, 2004], [2002, 2004]. We consider all 3-year windows starting between 2005 and 2014 for testing the predictions.
Table 4.
Top 10 predicted areas for inactive to active transition for professor Jussara Marques de Almeida, using both models, and actual top 10 by RCA. Time windows are set as in the first experimental setup.
Table 5.
Top 10 predicted areas for inactive to active transition for professor Renato Martins Assunção, using both models, and actual top 10 by RCA. Time windows are set as in the first experimental setup.
Fig 11.
Backbone with α = 0.20 of the space created by the embedding model for the intermediate areas, for different time intervals. The color of the edge represents whether it is inter-group (red) or intra-group (black). The color of the vertex represents the associated macro area.