Edge embedding with line graph enhancement

Abstract

Network data plays a crucial role in various real-world applications, as connections between entities can be represented and analyzed through graphs. These include diverse types such as social, information, and technical networks. However, the complex topologies of these networks present challenges in converting graph data into machine-readable vector formats. Existing models such as Graph Neural Network, Graph Attention Network, and node2vec have made strides in graph embeddings. Particularly for edge-related tasks, models like node2vec often resort to indirect methods like node concatenation for vector representation of edges, aiding in tasks like link sign prediction. In this paper, we introduce LineDi2vec, an innovative approach that enhances node2vec’s embedding method using a line graph. The proposed LineDi2vec not only generalizes the original graphs, transforming the relationships between edges and nodes but also maintains the original graphs’ topological integrity for effective node embedding by node2vec. We evaluated LineDi2vec on four real-world datasets, focusing on link prediction, link sign prediction and link direction prediction. The results demonstrate its superior performance over traditional node concatenation methods.