Beyond spatial neighbors: Utilizing multivariate transfer entropy for interpretable graph-based spatio–temporal forecasting

Abstract

Spatio–temporal forecasting is a challenging task that requires modeling complex interactions between multiple time series. While graph-based models have emerged as compelling tools for this task, their effectiveness heavily depends on the underlying graph structure that captures spatial dependencies but ignores the temporal relationships. To address this challenge, we propose the Multivariate Transfer Entropy-Multivariate Time series forecasting with Graph Neural Networks (MTE-MTGNN), a hybrid approach that combines statistical and deep learning methods. MTE-MTGNN introduces an interpretable graph construction layer founded on Multivariate Transfer Entropy, which effectively captures both spatial and temporal dependencies in the data. Empirical evaluations across five benchmark datasets demonstrate the superiority of our proposed approach in terms of predictive accuracy. The model shows particular strength in few-shot scenarios where traditional forecasting approaches typically struggle, achieving performance improvements of up to 3% on the RRSE metric in the exchange rate dataset and up to 4% on the correlation metric in the Hungarian Chickenpox dataset compared to state-of-the-art baselines. The findings witnessed across different experiments translate into significant practical benefits for real-world engineering applications and different domains.