Hybrid-hop multi-scale Line graph attention network for link prediction

Accurate link prediction plays a critical role in optimizing energy consumption and operational efficiency. In the context of power systems and wireless networks, numerous machine learning techniques have been applied to solve link prediction challenges from a graph-based perspective. However, these techniques often analyze topology features within a single scale from the extracted fixed-hop subgraphs of the whole network topology, which leads to overlooking latent configurations in different scopes of the existing graph systems. Subgraphs with different ranges may contain diverse numbers of potential correlations, typically represented as edges in the graph, which can influence the existence of links. Additionally, when subgraph range becomes excessively large, link prediction methods based on graph classification can result in the averaging of correlation information between nodes. Consequently, relying exclusively on a single scale for link existence prediction may lead to missed detections or incorrect predictions. To alleviate this limitation, this paper proposes a novel link prediction algorithm: Hybrid-hop multi-scale attention line graph network (HALT). HALT constructs hybrid-hop subgraphs for target nodes by applying a multi-hop segmentation approach during subgraph creation. These subgraphs are then converted into line graphs, transforming the problem from graph classification to graph node classification, which simplifies the computation process and reduces the information averaging issue. Finally, HALT employs a multi-scale attention neural network on the transformed line graph to capture potential correlations between links (i.e., between nodes in the line graph) across different ranges, ensuring more accurate link prediction. Experimental results demonstrate that the proposed HALT algorithm outperforms traditional methods, validating the effectiveness of the hybrid-hop subgraphs and multi-scale attention mechanism.