When exploring a data stream it is natural to ask how to relate current stream snapshots to past snapshots. Depending on the data semantics and the task at hand different interpretations are possible. For example, in the case of microblog data (like Twitter) making sense of conversations and discussions related to a particular topic may entice users to join the discussion. For data analysts, a usual task is to discern how tweets-information-patterns spread with the possible goal of intuitively explaining their findings. In monitoring traffic scenarios, teasing out those communication patterns that deviate from a considered normal behavior can be used as proxies for intrusion detection. In general social networks, identifying influential nodes in a “volatile” graph stream is of considerable interest. We report here a useful approach to identify trends and exceptional nodes in a graph stream. The fundamental idea is to view a graph stream as a collection of “elementary” time-stamped
events whose aggregation through time generates “salient” patterns whose activity rate is incrementally maintained. We are able to isolate group “herding” and “straying” as peculiar behaviors that can be subject to both human and computer verification. We quantitatively estimate the overall behavior of the detected salient edges as a convex combination of their “herding” or “straying” tendencies and their firing rates and recency “profiles.” All our computations are accompanied by a visualization platform that integrates dynamic node link views of “recent” subgraphs with a tape view of their Top-K edge statistics (Figure 12.1). The approach discussed here has been coupled with a degree-of-interest (DoI)-based exploration system  to provide the user with the functionality to take a closer look at particular keywords of interest identified with our novel approach. Currently, such DoI-based systems are not equipped to operate on the graph streaming setting proposed here.