A lightweight Spatial-Temporal Graph Neural Network for Long-term Time Series Forecasting

TL;DR

Lite-STGNN addresses long-term multivariate forecasting by uniting a decomposition-based temporal backbone with a lightweight, learnable sparse spatial module. The low-rank adjacency with Top-K sparsification enables scalable spatial reasoning that acts as a conservative residual to a strong linear baseline, yielding state-of-the-art results up to 720 steps while maintaining low parameter counts and fast training. Ablation studies confirm substantial gains from spatial modules and sparsity, and learned graphs align with domain structures, offering interpretability. The approach demonstrates a practical balance of accuracy, efficiency, and deployability across energy, finance, traffic, and weather domains.

Abstract

We propose Lite-STGNN, a lightweight spatial-temporal graph neural network for long-term multivariate forecasting that integrates decomposition-based temporal modeling with learnable sparse graph structure. The temporal module applies trend-seasonal decomposition, while the spatial module performs message passing with low-rank Top-$K$ adjacency learning and conservative horizon-wise gating, enabling spatial corrections that enhance a strong linear baseline. Lite-STGNN achieves state-of-the-art accuracy on four benchmark datasets for horizons up to 720 steps, while being parameter-efficient and substantially faster to train than transformer-based methods. Ablation studies show that the spatial module yields 4.6% improvement over the temporal baseline, Top-$K$ enhances locality by 3.3%, and learned adjacency matrices reveal domain-specific interaction dynamics. Lite-STGNN thus offers a compact, interpretable, and efficient framework for long-term multivariate time series forecasting.

Figures (6)

• Figure 1: Lite-STGNN architecture. A decomposition-based temporal module is combined with a lightweight spatial module that learns a low-rank Top-$K$ sparse adjacency to capture cross-variable dynamics. Conservative gating ($\sigma$), element-wise modulation ($\odot$), and residual addition ($+$) fuse spatial corrections with the temporal baseline.
• Figure 2: Per-horizon MSE over horizons 96/192/336/720 (mean over 3 runs). Lite-STGNN remains among the top performers, especially on Electricity and Exchange (Table \ref{['tab:main_results']}).
• Figure 3: Learned adjacency matrices (normalized weights) at horizon H720. Left: full connectivity; right: top 5% strongest edges.
• Figure 4: Representative forecast samples at H720. We observe similar results for other samples, confirming that these cases are representative.
• Figure 5: Ablation on Electricity (mean over 3 runs): spatial module improves MAE by 4.6% over temporal-only; rank=16 and Top-$K$=10 perform best and remain stable across horizons.