beyond models forecasting complex network strongprocessesstrong directly.pdfVIP

Beyond Models: Forecasting Complex Network Processes Directly from Data Bruno Ribeiro Minh X. Hoang Ambuj K. Singh Carnegie Mellon University University of California, University of California, Pittsburgh, PA, USA Santa Barbara, CA, USA Santa Barbara, CA, USA ribeiro@ mhoang@ ambuj@ ABSTRACT networks [12]; individuals also have evolving social inter- Complex network phenomena – such as information cascades ests [19, 24]. in online social networks – are hard to fully observe, model, In our work we focus on forecasting statistics of complex and forecast. In forecasting, a recent trend has been to forgo network processes, such as the distribution of sizes of an epi- the use of parsimonious models in favor of models with in- demic process over a network. Recently, the availability of creasingly large degrees of freedom that are trained to learn large-sample historical data – that details the evolution of the behavior of a process from historical data. Extrapolat- similar processes in the past – has started a trend of forgo- ing this trend into the future, eventually we would renounce ing parsimonious models in favor of models with increasingly models all together. But is it possible to forecast the evo- large degrees of freedom that are trained over this historical lution of a complex stochastic process directly from the data data; these include latent Markovian infection models [24], without a model? In this work we