ADAGE: A Framework for Generating Adaptable Intervals from Streaming Edges

We study the problem of determining the proper aggregation granularity for time-evolving network data when edges are added to the network in a streaming fashion. Time-evolving (a.k.a. longitudinal or dynamic) networks are often used to study topics such as change detection, evolution of communities, or network growth. However, aggregation lengths are often somewhat arbitrary, chosen based on intuition or convenience. The same network may be aggregated per-day in one study, per-week in another study, and per-month in yet another. It is unclear whether these interval lengths are appropriate for the tasks being considered. We describe a novel algorithmic framework, called ADAGE, which systematically detects the appropriate variable-length intervals based on convergence of a network statistics (e.g., exponent of the degree distribution) or performance on a graphmining task (e.g., true positive rate on belief propagation, BP). We apply ADAGE to 11 di↵erent network statistics and applications on 9 datasets from disparate domains. We observe that certain statistics consistently produce intervals suitable for various applications e.g., aggregation lengths based on the exponent of the degree distribution are suitable for BP. In addition, we present 2 case studies one on BP and another on network-similarity which demonstrate the usefulness of ADAGE in practice. We observe that in some applications (e.g., BP for malware detection) shorter aggregation lengths produce better performance.