思路:
首先,想在无标记的图结构中获得内在语义和结构属性?
问题:How to design an unsupervised learning task over the graph for pre-training the GNN model?
措施:p(G;θ):model the likelihood over this graph by this GNN, and representing how the nodes in G are attributed and connected
措施:最大化p(G;θ)
问题:how to model the conditional probality Pθ?
措施:dependency-aware factorization mechnism
措施:属性生成+边生成
问题:如何有效的优化属性生成和边生成任务?
措施:每个节点分为两部分,属性生成节点和边生成节点,得到属性生成和边生成的输出嵌入,计算属性生成和边生成的损失Loss
实验:
Different transfer settings.—time+filed+time/filed
结论:
the proposed generative pre-training strategy enables the GNN model to capture the generic structural and semantic knowledge of the input graph, which can be used to fine-tune on the unseen part of the graph data.
实验:
Ablation studies on pre-training tasks.—属性生成和边生成
结论:
the GPT-GNN framework benefits differently from attribute and edge generations on different datasets. However, combining the two pre-training tasks together produces the best performance on both cases.
the proposed graph generation tasks can give informative self-supervision for GNN pre-training
实验:
Ablation studies on the node separation and adaptive queue.
结论:
This demonstrates the significance of this node separation design in avoiding attribute information leakage.
This indicates that adding more negative samples by using the adaptive queue is indeed helpful to the pre-training framework.