Recent advancements in Graph Neural Networks (GNNs) have focused on improving transductive learning and representation learning. Researchers have introduced Training-Free Graph Neural Networks (TFGNNs) that utilize labels as features to enhance learning efficiency. Additionally, methods leveraging task structures have been proposed to improve the identifiability of neural network representations. New approaches also emphasize disentangled structural and featural representations for task-agnostic graph valuation. These innovations aim to address limitations in current GNN calculations and enhance the fairness of graph neural networks through techniques such as disentangling, amplifying, and debiasing. Furthermore, GNNRAI, a new model, integrates graphs to represent relationships among modality features, facilitating the encoding of biological knowledge in graph topology.