HP-JEPA Enhances Graph Representation Learning with Multi-Resolution

Graph self-supervised learning is crucial for extracting transferable representations from vast amounts of unlabeled graph data. While Joint-Embedding Predictive Architectures (JEPAs) have shown promise by predicting masked targets in latent space without explicit negative-pair construction, existing graph JEPAs typically rely on a single, predefined graph partition. This limitation biases the learned representations towards one structural granularity, hindering their ability to capture diverse patterns across different scales within a graph. To overcome this, the research proposes HP-JEPA, a hierarchical partitioning framework. HP-JEPA organizes each graph into an ordered bank of coarse-to-fine partition resolutions. It then performs context-target latent prediction independently at each resolution using an online encoder, a target encoder, and a latent predictor. The resulting resolution-specific graph representations are subsequently integrated, allowing downstream models to combine complementary local, regional, and global structural information. Experiments across various graph classification and regression benchmarks demonstrate that HP-JEPA consistently outperforms fixed-resolution Graph-JEPA baselines, particularly showing improved accuracy across different graph sizes.