Neural Additive Vector Autoregression Models for Causal Discovery in Time Series

Causal structure discovery in complex dynamical systems is an important challenge for many scientific domains. Although data from (interventional) experiments usually limited, large amounts of observational time series sets are available. Current methods that learn causal often assume linear relationships. Hence, they may fail realistic settings contain nonlinear relations between the variables. We propose Neural Additive Vector Autoregression (NAVAR) models, a neural approach to learning can discover train deep networks extract (additive) Granger influences evolution multi-variate series. The method achieves state-of-the-art results on various benchmark discovery, while providing clear interpretations mapped relations.