Label Distribution Learning on Subgraphs

Subgraph learning is one of the fundamental tasks in graph machine learning. Existing subgraph learning paradigms, including single-label learning and multi-label learning, rely on discrete supervision. In this study, we introduce label distribution learning (LDL) on subgraphs for the first time, enabling sub-graph learning with continuous-valued labels. The proposed task enhances the adaptability of subgraph learning to ambiguous data and limited annotation resources in real-world scenarios. Furthermore, we propose a general graph pooling architecture that is highly expressive, making it well-suited for label distribution learning, which is sensitive to input graph structure.