Learning Graph Neural Networks with Positive and Unlabeled Nodes

Graph neural networks (GNNs) are important tools for transductive learning tasks, such as node classification in graphs, due to their expressive power capturing complex interdependency between nodes. To enable graph network learning, existing works typically assume that labeled nodes, from two or multiple classes, provided, so a discriminative classifier can be learned the data. In reality, this assumption might too restrictive applications, users may only provide labels of interest single class small number addition, most GNN models aggregate information short distances (e.g., 1-hop neighbors) each round, and fail capture long distance relationship graphs. paper, we propose novel framework, long-short aggregation (LSDAN), overcome these limitations. By generating graphs at different levels, based on adjacency matrix, develop attention model The direct neighbors captured via short-distance mechanism, with by mechanism. Two risk estimators further employed long-short-distance networks, PU loss is back-propagated learning. Experimental results real-world datasets demonstrate effectiveness our algorithm.