In each case, we have sorted the proximity links according to their weight in descending order and calculated the confusion matrix after including more and more proximity links.
Again, we see that call links can be captured more efficiently by proximity links than Facebook interactions.
In the Appendix, we include further evaluation of how well proximity links can predict those of communication links.
(c) Fraction of social links recovered by the strongest ambient (dashed) and intimate (solid) physical proximity links: calls (green) and Facebook (blue).
It is important to realize that a majority of the proximity links (both ambient and intimate) are due to the co-location of students attending the same classes, which explains the low correlations seen in the data.
Here, the loss in the voice call graph is 9 %, and the introduction of the spatio-temporal proximity links increases the number of nodes weakly connected (more 3%% of nodes w.r.t (mathcal {D}_{textit {cs}})).
The most striking observation is that once around 5,000 of the strongest proximity links are considered (≈7% and 24% of links in the ambient and intimate networks, respectively), the contribution of additional links is comparably small (even negligible in case of call links).
Assuming that calls and Facebook activity (i.e., online communication) correspond to strong social ties, the capability of proximity links to predict those ties can be assessed by calculating the fraction of links in the former two networks that are covered by the strongest physical links.
Note that even if all physical proximity links are considered, we are not able to account for some of the links observed in the digital communication networks, which explains the absence of some precision-recall and ROC values in the interval ([0, 1]).
Given this proximity, linked physicians become progressively similar because of the induced homophily guiding their choices [ 24].
