@MASTERSTHESIS\{IMM2013-06615,
    author       = "A. Marcu",
    title        = "Predictability in Temporal Networks",
    year         = "2013",
    school       = "Technical University of Denmark, Department of Applied Mathematics and Computer Science / {DTU} Co",
    address      = "Matematiktorvet, Building 303B, {DK-}2800 Kgs. Lyngby, Denmark, compute@compute.dtu.dk",
    type         = "",
    note         = "{DTU} supervisors: Jakob Eg Larsen, jaeg@dtu.dk, and Sune Lehmann J{\o}rgensen, {DTU} Compute",
    url          = "http://www.compute.dtu.dk/English.aspx",
    abstract     = "With the increase in availability of temporal datasets collected from complex networks come new possibilities for studying the dynamic patterns formed by the interactions of such networks. Meaningful networks can be observed anywhere in day-to-day life: in phone-calls and daily social interactions; in public transportation, in technology and in nature: in interactions between species or between proteins. Having temporal data about such systems allows for a temporal networks representation.
While the link prediction problem already has developed well-established methods for predicting future interactions by analyzing a network’s intrinsic features, it predates the concept of temporal networks and only assumes a static network (a single state of the system), only being able to predict a single future state, of unknown temporal limits. When temporal data is available the expectations become higher, the occurrence of a new interaction has to be more precisely delimited in time, and more than a single state of the network has to be taken into account. Not much literature currently covers the prediction problem for temporal networks, and what exists is focused on certain domains and very specific approaches.
This thesis looks at the prediction problem for temporal networks from a broader angle, aiming to identify general goals for each stage of the problem, we propose an experimental framework for solving the prediction problem for temporal networks. The robustness of the framework is tested with an implementation aimed to obtain results from a temporal network of face-to-face interactions.
The results are collected from multiple experiments aimed to explore the parameter space, and are validated using state-of-the-art measures for predictive performance. These results will demonstrate that, although the specific methods were relatively simple, their implementation within the proposed framework brought relatively good results. The proposed framework is just a first step at generalizing a large problem, and the directions for further development are many: the framework could be optimized for specific domains or, by contrast, improved to provide more possibilities while keeping its generality."
}