1. intro

self-supervised pre-train + few-shot on down-tasks

之前的图生成网络VGAE等，不适合做pre-train：

1. 没有对特征进行生成；
2. 不适用于大规模图；

frame-work:

contribution:

1. decompose the attributed graph generation into 2 parts, namely attribution and edge generation and the 目标函数： 两者共同的极大似然
2. 高效的框架；

2.

和以往的图生成不同，GPT-GNN的预训练有着a harder graph task and thus can guide the model to learn more complex semantics and structure of the input graph

Our work is different with Strategies for Pre-training GNN[STRATEGIES FOR PRE-TRAINING GRAPH NEURAL NETWORKS ] as our goal is to pre-train GNNs over a single (large-scale) graph and conduct the node-level transfer.