Studies in Early Modern Social Networks, 1400-1750

Abstract

Recent developments in digital preservation have resulted in the creation of the “Six Degrees of Francis Bacon” (SDFB) network dataset from a portion of the Oxford Dictionary of National Bi- ography. This is one of the largest open-source historical social network datasets to date. At the same time, new advances in network science offer computationally efficient ways of characterizing large, complex networks. We apply these recently developed methods to the study of historical social networks. We produce three studies on the SDFB network. First, we apply multiplex PageRank to characterize networks with multiple layers of relation- ships as well as networks with multiple timescales. We validate this approach using recently de- veloped information-theoretic measures including the Jensen-Shannon distance. We find that ap- plying temporal multilayer PageRank appears to smooth out the independent popularity scores of timescale layers in a graph, which we argue creates a more realistic model of the change in an ac- tor’s influence in the network over time. Ranking on networks with multiple relationship layers contradicts historical wisdom, lending credence to figures such as royal tutors and bridesmaids with connections to highly influential people but few connections overall. Second, we use Robin Forman’s adaptation of Ricci curvature, an edge-based measure of “bridg- ing” behavior for graphs, to examine key relationships in the network. We demonstrate that this measure is computationally efficient in creating graph “backbones.” Applying Forman-Ricci cur- vature highlights the role of writers, philosophers, and humanists as go-betweens between royalty and communities of intellectuals, though we find these go-betweens must be closely allied to roy- alty to gain intermediary status. We propose a novel adaptation of Forman-Ricci curvature for multiplex networks by applying the recently introduced notion of a hypergraph to multiplex net- works. After testing this approach on a smaller dataset of Renaissance Florentine families, we find that this new measure of hyperedge curvature amplifies the role of women as conduits for influence in the SDFB network. Finally, we use supervised link classification to predict potential relationships on the SDFB net- work. We develop classifiers that integrate features from metadata in the SDFB dataset as well as from the underlying graph topology which predict relationships well (AUC = .94 in the best case) on a partitioned set of test data. Further, we show that many of the highest-probability false positives are actually historically attested at least one-third of the time, suggesting that link pre- diction could be a powerful method not only to infer missing relationships and improve database construction, but also to shed light on ties previously unknown to historians.