On mining temporal patterns in dynamic graphs, and other unrelated problems

Abstract

Given a social network with dynamic interactions, how can we discover frequent interactions between groups of entities? What are the temporal patterns exhibited by these interactions? Which entities interact frequently with each other before, during, or after others have stopped or started? Such dynamic-network datasets are becoming prevailing, as modern data-gathering capabilities allow to record not only a static view of the network structure, but also detailed activity of the network entities and interactions along the network edges. Analysis of dynamic networks has applications in telecommunication networks, social network analysis, computational biology, and more. We study the problem of mining interactions in dynamic graphs. We assume that these interactions are not instantaneous, but more naturally, each interaction has a duration. We solve the problem of mining dynamic graphs by establishing a novel connection with the problem of mining event-interval sequences, and adapting methods from the latter domain. We apply the proposed methods to a real-world social network and to dynamic graphs from the field of sports. In addition, having established the aforementioned equivalence between the two pattern-mining settings, we proceed to describe how other graph-related problems, such as prediction, learning, and summarization, can be solved by applying out-of-the-box algorithms devised for event-interval sequences. In light of these results, we conjecture that there may be further connections between the two research domains, and the two communities should work closer to share goals and methodology.