Effects of Node Centrality Measures for Classification Tasks using GNNs
Keywords: graph neural networks, road type prediction, network embedding, graph machine learning
Abstract: This study explores the impact of feature selection, particularly node centrality measures, on road type classification within a road network graph using Graph Neural Networks (GNNs) and traditional machine learning models. By training six models on three distinct feature sets—primary road characteristics (S1), centrality measures (S2), and a combined feature set (S3)—we analyze how different feature representations affect model accuracy in distinguishing road types. The GraphSAGE model using S1 achieved the highest test accuracy (0.89), indicating that primary road characteristics are highly effective for classification, whereas the Random Forest model performed worst on the same set, achieving only 0.17 accuracy. Visualized embeddings from S1 models reveal effective clustering by road type for models like GraphSAGE, particularly for residential and tertiary roads, underscoring the model’s capability to capture nuanced structural relationships. These findings indicate that feature selection, especially the inclusion of relevant node centrality measures, plays a crucial role in enhancing classification, though further improvement may require hybrid models or additional contextual data sources to address limitations in differentiating road types with overlapping attributes.
Submission Number: 43
Loading