[1. Introduction]

1. transductive learning: node & edge가 training~prediction phase에서 그대로 남아있음 (access 가능). unseen node와 edge 처리 못함
2. inductive learning: 그래프 일부에서만 학습하고, 그 모델을 unseen node or deges를 예측하는 데 사용

• Paradigm: spectral architecture on undirected graph in transductive settings → spatial ones that can apply inductively to arbitrary graphs

{contribution}

1. discuss some theoretical tools to better visualize the operations performed by spatial GNNs
2. analyze the inner workings of the SOTA spatial architecture,s specifically aggregation-based GNNs
3. propose simple technique: Node masking

[2. A brief history of GNNs]

[2.1 Spectral GNNs] - message passing

• learns spectral filters regard to Laplcian → lack of scalibility
• ChebNet, GCN

[2.2 Spatial GNNs]

• defines the convolution operation directly on the structure of the graph
• sampling-based: GraphSAGE
• aggregation-based(Generally SOTA): GAT, GIN(maximally powerful)

[4. Analysis of Aggreagation-based GNNs]

[5. Node masking]

• node masking of GAT and GIN (respectively)

[5.2 Node masking vs Dropout techniques]

• dropedge: randomly drops a subset of edges from the graph in every training epoch
• in our experiment, node masking이 dropedge보다 효과 좋음

[6.3 Experimental setting]

• pytorch, 2-layer GIN
• simple GAT (SGAT)
• if node mask active → in every training epoch, randomly sample a Bernolli select function over the nodes in the training set
• transductive- make entire graph
• inductive- make avalibale at training time the graph G formed by nodes in the train set + edges among them