LG-GNNs: Enhancing Node Classification via Line Graph Transformations for Higher-Order Structural Learning

Graph Neural Networks (GNNs) have demonstrated remarkable success in learning robust representations from complex graph-structured data. However, standard GNNs architectures often fail to capture higher-order structural information effectively, resulting in suboptimal performance on tasks that require a nuanced understanding of intricate node dependencies. In this work, we present a novel architectural innovation—the Line Graph GNN (LG-GNNs)—which integrates the line graph transformation into the traditional GNNs framework. By explicitly modeling higher-order interactions among edges, our approach enriches node embeddings with additional structural context. We evaluate LG-GNNs on a variety of benchmark datasets that capture both simple and complex network structures, comparing its performance against conventional GNNs architectures. Experimental results demonstrate that LG-GNNs consistently outperforms baseline models in terms of classification accuracy while maintaining competitive efficiency. Motivated by the shortcomings of standard GNNs in capturing complex relational dependencies in real-world networks, our method delivers a more expressive representation that enhances predictive performance across diverse applications. To encourage reproducibility and further research in the field, we have made our implementation and experimental results publicly available.