Discovery of anomalous behaviour in temporal networks

In this work we consider the problem of detecting anomalous behaviour and present a novel approach that allows 'behaviour' to be classified as either to be normal or abnormal by checking the p-value associated with the occurrence of the behaviour which is modelled following a binomial distribution within a discrete time model. We investigate the problem of detecting anomalous behaviour by looking at how communication evolves over time in a social network graph. Under the assumption that some nodes of the network could be labelled qualitatively, we present a novel approach that allows us to infer a subset of nodes of the social network which might share the same qualitative connotation. In other words, assuming one or more members belong to some criminal organisation, we wish to investigate how many other persons belong to the same organisation. We have tested our method in two datasets, VAST2008 and a Twitter Dataset (data collected in 2012), with encouraging results.