The continued proliferation of timestamped network data demands increasing sophistication in the analysis of that data. In particular, the literature amply demonstrates that the choice of temporal resolution has a profound impact on the solutions produced by many different methods in this domain -- answers differ when data is viewed second-by-second as opposed to week-by-week. Additionally, research also shows quite clearly that the rates at which network events happen are not constant -- some times are "faster" or "slower" than others, and these variations are not necessarily predictable. Given the above, it is clear that there must be problem settings in which no fixed choice of temporal resolution will produce ideal results.

In this work, we explore several approaches for utilizing non-uniform temporal resolution to improve upon existing analysis. When events of interest are localized to certain portions of the timeline and the network, we demonstrate a fast method that prunes the portions of the timeline that are not relevant and then uses a novel temporal variant of locality-sensitive hashing (LSH) that is effective for identifying bursty low-conductance communities in the network. When consideration of the full timeline is necessary, we demonstrate techniques for finding accordion transformations -- order-preserving compressions and expansions -- of the timeline that improve the results of known methods for link prediction, cascade size prediction, interpolation of missing event data, and other tasks.