Dyn-GWN: Time-Series Forecasting using Time-varying Graphs with Applications to Finance and Traffic Prediction

Abstract

Spatio-temporal modeling is an essential lens to understand many real-world phenomena from traffic to finance. There has been exciting work that explores spatio-temporal modeling with temporal graph convolutional networks. Often these methods assume that the spatial structure is static. We propose a new model Dyn-GWN for spatio-temporal learning from time-varying graphs. Our model relies on a novel module called the Tensor Graph Convolutional Module (TGCM), which captures dynamic trends in graphs effectively in the time-varying graph representations. This module has two components: (i) it applies temporal dilated convolutions both on the time-varying graph adjacency space and the time-varying features. (ii) it aggregates the higher-level latent representations from both time-varying components through a proposed layer TGCL. Experiments demonstrate the efficacy of these model across time-series data from finance and traffic domains. Dyn-GWN can give up to better out-of-sample performance than prior methods that learn from time-varying graphs, e.g., EvolveGCN and TM-GCN. Interestingly, Dyn-GWN can be ∼ 300 × faster than EvolveGCN, which is the more competitive baseline from state-of-the-art models that cater to time-varying graphs.