Fig 1.
Mean average error over 1000 training epochs.
Mean average error of the neural network, i.e. the average difference between the true and predicted author H-Index, for each epoch. After around 400 epochs, the network starts to overfit to the training data, as the training error rate continues to go down, while the validation error rate (shown as a dotted line) remains roughly constant.
Fig 2.
Visualisation of co-authorship network for DCM research.
Colours show country. Author node size shows number of papers published. Authors shown as lone nodes do not have any collaborations with the other “prolific” selected authors.
Table 1.
Top 10 authors, by various network measures.
Fig 3.
Distributions of various network metrics with H-Index.
Average DCM papers of co-authors is the total number of papers an author’s co-authors have published. Connected institutes is the number of distinct research institutes in which the author had co-authors. Connections are the total number of an author’s co-authors. Second order connections are the total number of the co-authors’ co-authors.
Table 2.
Statistical data about authors in the co-authorship network.
Fig 4.
True and predicted values for H-Index.
Predictions up to an H-Index of 50 are shown.
Fig 5.
KDE plot of h-index prediction error.
The prediction error is the difference between the true and predicted value. Thus, a negative number shows the network under-predicted, a positive value showed the network over-predicted.
Fig 6.
Variation in predicted H-Index with author connectedness statistics.
The bottom axis shows the standard deviations from the mean; the top axis shows the actual number.