Mining Dynamic Networks: The Importance of Pre-processing on Downstream Analytics

Dynamic networks are becoming ubiquitous and the analysis of these is becoming increasingly important to better understand the processes by which they evolve over time. Various recent work in this space have looked at how to detect communities, how to model adding/removing nodes and edges, and how to model how nodes change roles over time. In this paper, we look at a different aspect of dynamic networks: assuming we can identify and track a community over time, can we predict larger evolutionary events such as whether it is about to merge with another community or split or dissolve. Further, can we predict at a node-level whether nodes are about to leave the community. We here provide an initial exploration of how different settings for network extraction from observed time-stamped links impact community detection and performance of machine learning. We show on two data sets that changing these parameters have drastic impact on the consistency and stability of communities found.