We study the problem of detecting top-k events from digital interaction records (e.g, emails, tweets). We first introduce interaction meta-graph, which connects associated interactions. Then, we define an event to be a subset of interactions that (i) are topically and temporally close and (ii) correspond to a tree capturing information flow. Finding the best event leads to one variant of prize-collecting Steiner-tree problem, for which three methods are proposed. Finding the top-k events maps to maximum k-coverage problem. Evaluation on real datasets shows our methods detect meaningful events.
