Predicting traffic incident risks at granular spatiotemporal levels is challenging. The datasets predominantly feature zero values, indicating no incidents, with sporadic high-risk values for severe incidents. Notably, a majority of current models, especially deep learning methods, focus solely on estimating risk values, overlooking the uncertainties arising from the inherently unpredictable nature of incidents. To tackle this challenge, we introduce the Spatiotemporal Zero-Inflated Tweedie Graph Neural Networks (STZITD-GNNs). Our model merges the reliability of traditional statistical models with the flexibility of graph neural networks, aiming to precisely quantify uncertainties associated with road-level traffic incident risks. This model strategically employs a compound model from the Tweedie family, as a Poisson distribution to model risk frequency and a Gamma distribution to account for incident severity. Furthermore, a zero-inflated component helps to identify the non-incident risk scenarios. As a result, the STZITD-GNNs effectively capture the dataset's skewed distribution, placing emphasis on infrequent but impactful severe incidents. Empirical tests using real-world traffic data from London, UK, demonstrate that our model excels beyond current benchmarks. The forte of STZITD-GNN resides not only in its accuracy but also in its adeptness at curtailing uncertainties, delivering robust predictions over short (7 days) and extended (14 days) timeframes.

翻译：在精细时空层面上预测交通事件风险极具挑战性。数据集中绝大部分取值为零，表示无事件发生，仅偶有高风险值对应严重事件。值得注意的是，现有大多数模型（尤其是深度学习方法）仅专注于估计风险值，忽视了由事件固有不可预测性所产生的不确定性。为解决这一难题，我们提出了时空零膨胀Tweedie图神经网络（STZITD-GNNs）。该模型融合了传统统计模型的可靠性与图神经网络的灵活性，旨在精确量化道路级交通事件风险的相关不确定性。我们策略性地采用Tweedie分布族中的复合模型，以泊松分布建模事件发生频率，以伽马分布描述事件严重程度。此外，零膨胀组件有助于识别非事件风险场景。因此，STZITD-GNNs能够有效捕捉数据集的偏态分布，重点关注罕见但具有显著影响的严重事件。基于英国伦敦真实交通数据的实证测试表明，该模型超越现有基准方法。STZITD-GNN的优势不仅体现在预测精度上，还在于其有效降低不确定性的能力，可在短周期（7天）和长周期（14天）内提供稳健预测。