Using non-negative factorization of time series of graphs for learning from an event-actor network

While non-negative factorization is a popular tool for analyzing non-negative data, current model selection methods can perform poorly for non-negative factorization when dealing with stochastic data. We develop model selection techniques that can be used to augment existing non-negative factorization algorithms, illustrating the performance of our algorithms via the application to problems of inference on time series of graphs. We motivate our approach with singular value decomposition, and illustrate our framework through numerical experiments using real and simulated data.