Spatio-Temporal Time Series Forecasting with Graph-Enhanced Embeddings

Abstract

Spatio-temporal time series forecasting is an important task in many areas such as traffic networks, environmental monitoring, and social networks. The goal is to predict future values of signals on the nodes of a graph using past observations. Current temporal encoding methods assume that temporal patterns are the same for all nodes. In practice, however, different nodes can show different temporal behavior depending on where they are in the graph and what their neighbors look like. In this work, we propose Graph-Enhanced Embeddings, a method that produces temporal embeddings based on the node's own signal and its neighbors' signals. The method uses learnable sinusoidal bases and a linear trend term together with a gating mechanism that controls which temporal bases are active for each node. We test the method on three benchmark datasets and show that it consistently outperforms standard baselines including fixed positional encodings, linear features, and Time2Vec.