Structural and Functional Discovery in Dynamic Networks with Non-negative Matrix Factorization

Due to advances in data collection technologies, it is becoming increasingly common to study time series of networks. An important research question is how to discover the underlying structure and dynamics in time-varying networked systems. In this work, we propose a new matrix factorization-based approach for community discovery and visual exploration within potentially weighted and directed network time-series. Next, we review and discuss this work in relation to popular approaches for addressing the key problems of community detection and visualization of time series of networks. There have been many important contributions for community detection in network time-series, extensively reviewed in [1, 2], from the fields of physics, computer science and statistics. The basic goal of community detection is to extract groups of nodes that feature relatively dense within group connectivity and sparser between group connections [3, 4]. A common strategy is to embed the graphs in low-dimensional latent spaces. For instance, [5] use latent variables to capture groups of papers that evolve similarly in citation network data.