About ablation studies on base GNNs about gpt-gnn HOT 2 CLOSED

In table 1, GPT-GNN on whole OAG under field transfer setting is .407±.107. (https://arxiv.org/pdf/2006.15437.pdf). 0.393 is for time + field transfer.

We treat GAE, GraphSage (unsp.), and Graph Infomax as pre-training paradigm, so they only provide unsupervised guidance to train the GNN model. The default GNN model is HGT.

To evaluate the different encoders, we have this Table 2. To handle large-scale graphs, we use the same sampling procedure as described (and implemented) similar to HGT. For GCN and GAT, we simply treat the sampled graph as homogeneous graph, while maintaining the first layer (feature adaptation) for a different types of nodes.

from gpt-gnn.

Thank you for your response!

from gpt-gnn.