PostDoc - Higher-Order Networks for the Analysis of Geo-Historical Traces
Network analysis - sequences mining - digital humanities - mobility traces - graphsalgorithms
Equipes : DUKe - Data User Knowledge
Etablissement : Ecole Polytechnique de l'Université de Nantes (Polytech Nantes)
Description du poste :
ContextNetworks analysis methods are now well established research tools in Digital Humanities. However, data used to built the networks sometimes correspond to sequences of events (movements in a city, hypertext links clicked by a website user, list of ports visited by a cargo ship, etc.). Recent works [1,2] challenge the aggregation of these sequence (Fig. 1a) as sums of independent exchanges between pairs of entities (weighted edges in a graph) (Fig. 1b). These aggregated models of networks implicitly correspond to memory-less model i.e. first order Markov chains (during a random walk, the next node visited depends only on the last visited node). These models can be used to predict future sequences  but produce poor results in many cases (Fig. 1c).
ProjectAs it is often the case in statistical modeling, the difficulty is to built good predictive models that are parsimonious enough to produce valuable information. In the context of geo-historical datasets, we will be interested in the comparison between higher-order models and models taking into account exogenous variables (e.g. the continent or type of port visited by a ship). The objective here is to evaluate the relevance of higher-order when compared to others network models such as multi-layered or labeled graphs. A second issue is the design of mining tools for higher-order networks like, for instance, the redefinition of "classic" tools of network analysis (e.g. graph clustering or centrality measures). We will focus on the impact of the choices used to built higher-order networks on the results of mining algorithms. A model recently proposed by Xu et al.  uses vertex-duplication to preserve sequential information. The authors suggest that conventional network analysis tools can be directly applied on this new network. However, one can assume that there is not a single way to duplicate the vertices to produce a good predictive model. One challenge is to quantify the impact of this bias. The "traditional" representation of networks as graphs is today common practice in digital humanities. This new field of network analysis is therefore a interdisciplinary challenge. It is important to evaluate the relevance of this new approach for researchers in Humanities.
Candidate profile & PositionThe candidate should have experience in Network Analysis and/or Sequential data analysis and prediction. Good coding skills (C++, python). Interest for and/or experience in interdisciplinary research (Digital Humanities) is a plus.
- approximately 2200 €/ month (net)
- flexible starting date
- Francois Queyroi : francois.queyroiatls2n.fr