I. Introduction
We study short term state prediction, the task of predicting the picture of a network at time , based on historical data. The directed network is composed of links, each link being characterized by a state at time . For many large scale networks such as energy, transport, or internet, a current challenge relies on the ability to scale up prediction performances based on complex spatiotemporal data. It is reached thanks to Multi-Input Multi-Output (MIMO) supervised regressions from K-NN [1] to NN [2] and M-SVR [3]. The main challenge of these approaches lies in the high-dimensionnality and dynamic of the system. From a supervised learning standpoint, this implies to either propose a dynamic compact representations of the individuals or to dynamically reduce inputs dimensions [4], [5], coping with relevance and redundancy [6]. This common objective is shared by many application fields [7] with various methods [8] and the interest in the ML community is still strong due to increasing dimensionality [9], constrained budget [10] or even security [11] constraints. A second challenge lies in the resilience to non recurrent network conditions such as traffic congestion. Hence, the research issue can be summarized as the selection of the best subset of critical links (inputs of the system) in spatiotemporal networks with following properties:
Best trade-off between dimensionality reduction and generalization performances,
Resilience to recurrent or non recurrent congested phases,
Understanding of the relationship between dynamics of the physical system and network's preditability.