Abstract
Analysing traffic data is an important task in the context of intelligent transportation systems within cities. Graph convolutional neural networks (GCNNs) have appeared to be an important tool for performing this task due to their promising performance in extracting spatial correlations in graphs. Nevertheless, the design of these networks and the effectiveness of the networks’ layers-or components-remains unclear and lacks intuition. This study aims to bridge this gap by conducting an ablation study that compares the impact of components that are commonly used in the literature, including spatial and temporal convolution layers within GCNN models performance. To achieve this goal, we utilize a baseline network wherein each layer can be replaced by different approaches, allowing the generation of results under various scenarios. Our focus is on resolving the problem of estimating missing values for road segments’ average speeds. We examine various GCNN components utilizing two real-world datasets to assess their effectiveness.