Panning for gold: using variograms to select useful connections in a temporal multigraph setting

As social networks continue growing in size and popularity, so does the amount of data that is publically available pertaining to the members of these networks, which results in a rapid growth of the types of connections that can potentially be established between members of the network. With this increasing amount of available information, the task of examining social networks becomes more exciting and more difficult at the same time. Researchers now have to filter the large field of potentially useful information to find the golden nuggets that will serve their needs, while ignoring data that would actually hamper their experiments. To address this issue, we propose a novel integration of a variogram-based system for selecting potentially informative relationships with a Gaussian Conditional Random Field (GCRF) model that is able to perform node attribute prediction in social networks that are both temporal and contain a wide variety of relationship types. By first evaluating the large pool of relationships via variograms, which can be done very quickly, we are able to fully utilize the power of the GCRF model, which becomes computationally expensive when the network size or the number of relationships examined grows. Our experiments show that we can use variograms to narrow the potential field of relationships drastically and consequently make the GCRF model run in a reasonable amount of time and remain accurate.