Multi-task Graph Neural Network for Optimizing the Structure Fairness

Graph neural networks, the mainstream paradigm of graph data mining, optimize the traditional feature-based node classification models with supplementing spatial topology. However, those isolated nodes not well connected to the whole graph are difficult to capture effective information through structural aggregation and sometimes even bring the negative local over-smoothing phenomenon, which is called structure fairness problem. To the best of our knowledge, current methods mainly focus on amending the network structure to improve the expressiveness with absence of the influence of the isolated parts. To facilitate this line of research, we innovatively propose a Multi-task Graph Neural Network for Optimizing the Structure Fairness (GNN-OSF). In GNN-OSF, nodes set is divided into diverse positions with a comprehensive investigation of the correlation between node position and accuracy in global topology. Besides, the link matrix is constructed to express the consistency of node labels, which expects isolated nodes to learn the same embedding and label when nodes share similar features. Afterward, the GNN-OSF network structure is explored by introducing the auxiliary link prediction task, where the task-shared and task-specific layer of diverse tasks are integrated with the auto-encoder architecture. Our comprehensive experiments demonstrate that GNN-OSF achieves superior node classification performance on both public benchmark and real-world industrial datasets, which effectively alleviates the structure unfairness of the isolated parts and leverages off-the -shelf models with the interaction of auxiliary tasks.