Selecting transfer entropy thresholds for influence network prediction

Determining the influence of participants within a network is important for a number of reasons, including maximizing advertising revenue and efficiency, behaviour prediction and control. Structural analysis has been widely applied to predict user influence within social systems, but this analysis requires the underlying network to be known. Furthermore, structural analysis generally assumes that links connecting users represent identical relationships, which does not account for possible differences stemming from trust or other relationship factors. Previous work has used transfer entropy measurements between time series as a form of influence measure in social systems, but has not considered the problem of identifying which possible links are significantly influential in relation to others. This work builds upon the existing work by using transfer entropy to predict a graph, in which links represent influence between two agents. The main contribution of this work is the proposal of methods for selecting a transfer entropy threshold value which determines whether a link should be considered influential. These methods could be used to either identify unknown links within a network or prune known links that do not actually represent a true influence relationship. The results presented show that this method generally allows the influence network to be predicted with precision and recall values of over 90% across three common theoretical network classes.