The COVID-19 pandemic has claimed millions of lives, spurring the development of diverse forecasting models. In this context, the true utility of complex spatio-temporal architectures versus simpler temporal baselines remains a subject of debate. Here, we show that structural sparsification of the input graph and temporal granularity are determining factors for the effectiveness of Graph Neural Networks (GNNs). By leveraging human mobility networks in Brazil and China, we address a conflicting scenario in the literature: while standard LSTMs suffice for smooth, monotonic cumulative trends, GNNs significantly outperform baselines when forecasting volatile daily case counts. We show that backbone extraction substantially enhances predictive stability and reduces predictive error by removing negligible connections. Our results indicate that incorporating spatial dependencies is essential for modeling complex dynamics. Specifically, GNN architectures such as GCRN and GCLSTM outperform the LSTM baseline (Nemenyi test, p < 0.05) on datasets from Brazil and China for daily case predictions. Lastly, we frame the problem as a binary classification task to better analyze the dependency between context sizes and prediction horizons.

翻译：COVID-19疫情已导致数百万人死亡，推动了多种预测模型的发展。在此背景下，复杂的时空架构相较于更简单的时间基线模型的实际效用仍存在争议。本研究表明，输入图的结构稀疏化和时间粒度是决定图神经网络（GNN）有效性的关键因素。通过利用巴西和中国的人类移动网络，我们解决了文献中的矛盾情景：当预测平滑单调的累积趋势时，标准LSTM模型已足够；而在预测波动的每日病例数时，GNN显著优于基线模型。我们证明，通过移除无关连接进行骨干网络提取可大幅增强预测稳定性，并降低预测误差。结果表明，纳入空间依赖关系对建模复杂动态过程至关重要。具体而言，GCRN与GCLSTM等GNN架构在巴西和中国的每日病例预测数据集上优于LSTM基线模型（Nemenyi检验，p < 0.05）。最后，我们将该问题转化为二分类任务，以更深入地分析上下文规模与预测时间跨度之间的依赖关系。