Graph neural networks (GNNs) have shown strong performance in applications like social networks and recommender systems, especially on large-scale graphs. However, training GNNs on such graphs suffers from the neighbor explosion problem, leading to high memory consumption. To tackle this, the "graph partitioning \( + \) local learning" framework splits the global graph into smaller subgraphs for independent training. Existing graph partitioning methods fall into node and edge partitioning. Node partitioning drops inter-subgraph edges, causing incomplete information, while edge partitioning, though structure-preserving, introduces two major challenges: (1) Existing edge partitioning methods often rely on node degrees to divide subgraphs, ignoring the actual role of nodes in information propagation, which may result in the loss of critical information paths and affect training accuracy; (2) during subgraph training, nodes cannot access all their original neighbors, limiting the model’s expressive capacity and reducing predictive performance. To overcome these issues, we propose HGP-IC, a model combining a refined partitioning strategy (HGP) with an information compensation (IC). HGP not only resolves memory constraints during GNNs training on large-scale graph data-enabling tasks originally designed for limited resources to execute smoothly but also mitigates the impact of node partitioning during training. IC restores lost neighbor information between subgraphs, enhancing representation and predictive performance. Experiments on seven public datasets confirm that HGP-IC significantly improves GNNs accuracy and demonstrates strong effectiveness on large-scale graphs.https://github.com/huang-1314-xu/HGP-IC.git.
