Temporal Graph MLP Mixer for Spatio-Temporal Forecasting
Muhammad Bilal, Luis Carretero Lopez
TL;DR
The paper introduces the Temporal Graph-MLP Mixer (T-GMM) to tackle spatiotemporal forecasting under substantial missing data by fusing node-level processing, patch-based subgraph encoding, and a 3D MLP-Mixer to jointly model temporal, spatial, and feature interactions. Through comparative analyses and targeted experiments, it demonstrates that low-inductive-bias architectures can achieve competitive performance given ample data and compute, while the T-GMM architecture provides robustness to long-range dependencies and missing data, albeit with memory efficiency challenges. The work advances understanding of inductive biases in spatiotemporal forecasting and outlines directions for reducing memory usage and improving generalization under non-stationary conditions. The reproducibility emphasis and open-source resources support broader adoption and further exploration of robust forecasting in sensor networks.
Abstract
Spatiotemporal forecasting is critical in applications such as traffic prediction, climate modeling, and environmental monitoring. However, the prevalence of missing data in real-world sensor networks significantly complicates this task. In this paper, we introduce the Temporal Graph MLP-Mixer (T-GMM), a novel architecture designed to address these challenges. The model combines node-level processing with patch-level subgraph encoding to capture localized spatial dependencies while leveraging a three-dimensional MLP-Mixer to handle temporal, spatial, and feature-based dependencies. Experiments on the AQI, ENGRAD, PV-US and METR-LA datasets demonstrate the model's ability to effectively forecast even in the presence of significant missing data. While not surpassing state-of-the-art models in all scenarios, the T-GMM exhibits strong learning capabilities, particularly in capturing long-range dependencies. These results highlight its potential for robust, scalable spatiotemporal forecasting.