We propose an architecture for multivariate time-series prediction that integrates a spatial-temporal graph neural network with a filtering module which filters the inverse correlation matrix into a sparse network structure. In contrast with existing sparsification methods adopted in graph neural networks, our model explicitly leverages time-series filtering to overcome the low signal-to-noise ratio typical of complex systems data. We present a set of experiments, where we predict future sales volume from a synthetic time-series sales volume dataset. The proposed spatial-temporal graph neural network displays superior performances to baseline approaches with no graphical information, fully connected, disconnected graphs, and unfiltered graphs, as well as the state-of-the-art spatial-temporal GNN. Comparison of the results with Diffusion Convolutional Recurrent Neural Network (DCRNN) suggests that, by combining a (inferior) GNN with graph sparsification and filtering, one can achieve comparable or better efficacy than the state-of-the-art in multivariate time-series regression.
CCS CONCEPTS• Computing methodologies → Machine learning approaches.