Traffic flow prediction is an essential component in intelligent transportation systems. Recently, there has been a notable trend in applying machine learning models, especially deep learning, for network-wide traffic prediction. However, existing studies have limitations on model interpretability, model generalization, and over-reliance on image data processing or fine-designed deep learning structures for extracting traffic attributes. This paper attempts to tackle these limitations by proposing a Bayesian clustering ensemble Gaussian process (BCEGP) model for network-wide traffic flow clustering and prediction. The model utilizes a subset-based Dirichlet process mixture (SDPM) model to conduct a hard clustering among input data; then, within each cluster, it adopts the Gaussian Process (GP) to learn the probability relationship between inputs and outputs. During the prediction phase, the model conducts a soft clustering of the input as weights, and makes predictions via a weighted average of GPs’ outputs. The merits of the BCEGP model include: (a) data with similar spatial–temporal patterns are clustered, which helps understand traffic dynamics in a non-Euclidean and non-graphical manner that enhances information extracting for model development; (b) GPs provide analytically trackable functions/gradients of predicted traffic flows with features and reveal variances of predicted traffic flow, enhancing model applicability and interpretability to some extent; (c) the model incorporates an ensemble learning framework that achieves great generalization performance as good as deep learning models; (d) the subset-based clustering and cluster-based GP learning are conducted parallelly, and thus vastly accelerate the training efficiency compared with conventional GPs (but slower than deep learning models). We test the performance of the proposed model based on both synthesized and real-world datasets. For comparison, several widely used machine learning and deep learning models are trained under the real-world dataset. The results demonstrate that the BCEGP model performs well in predictive accuracy, computational speed, and applicability, which can be a promising method for transportation problems.

交通流量预测是智能交通系统的重要组成部分。最近，将机器学习模型，尤其是深度学习应用于全网流量预测的趋势很明显。然而，现有研究在模型可解释性、模型泛化以及过度依赖图像数据处理或精心设计的深度学习结构来提取交通属性方面存在局限性。本文试图通过提出用于网络范围的交通流聚类和预测的贝叶斯聚类集成高斯过程 (BCEGP) 模型来解决这些限制。该模型利用基于子集的 Dirichlet 过程混合 (SDPM) 模型在输入数据之间进行硬聚类；然后，在每个集群中，它采用高斯过程（GP）来学习输入和输出之间的概率关系。在预测阶段，模型将输入作为权重进行软聚类，并通过 GP 输出的加权平均值进行预测。BCEGP 模型的优点包括：（a）具有相似时空模式的数据被聚类，这有助于以非欧几里德和非图形方式理解交通动态，增强模型开发的信息提取；(b) GP 提供具有特征的预测交通流的分析可跟踪函数/梯度，并揭示预测交通流的方差，在一定程度上增强模型的适用性和可解释性；(c) 该模型结合了一个集成学习框架，实现了与深度学习模型一样好的泛化性能；(d) 基于子集的聚类和基于聚类的 GP 学习是并行进行的，因此与传统 GP 相比大大提高了训练效率（但比深度学习模型慢）。我们基于合成数据集和真实数据集测试了所提出模型的性能。为了进行比较，几个广泛使用的机器学习和深度学习模型是在真实世界的数据集下训练的。结果表明，BCEGP 模型在预测精度、计算速度和适用性方面表现良好，是解决交通问题的一种很有前途的方法。我们基于合成数据集和真实数据集测试了所提出模型的性能。为了进行比较，几个广泛使用的机器学习和深度学习模型是在真实世界的数据集下训练的。结果表明，BCEGP 模型在预测精度、计算速度和适用性方面表现良好，是解决交通问题的一种很有前途的方法。我们基于合成数据集和真实数据集测试了所提出模型的性能。为了进行比较，几个广泛使用的机器学习和深度学习模型是在真实世界的数据集下训练的。结果表明，BCEGP 模型在预测精度、计算速度和适用性方面表现良好，是解决交通问题的一种很有前途的方法。