๐Paper Review
[paper review] Heterogeneous Graph Attention Network
Introduction
๊ธฐ์กด์ Graph Attention network๋ฅผ Heterogeneous graph์ ์ ์ฉํ๊ธฐ ์ํ ๋ฐฉ๋ฒ๋ก ์
๋๋ค.
None stectral(spatial) GNN์ด๊ธฐ ๋๋ฌธ์ transductive problem ๋ฟ๋ง ์๋๋ผ, Inductive problem์๋ ์ ์ฉํ ์ ์์ต๋๋ค.
ย
๋ณธ ๋
ผ๋ฌธ์ Contribution์ ๋ค์๊ณผ ๊ฐ์ต๋๋ค.
- attention mechanism์ Heterogeneous graph์ ์ ์ฉํ GNN ๋ฐฉ๋ฒ๋ก ์ ๋๋ค.
- aggregation ๊ณผ์ ์์ Node-level, Metapath(Semantic)-level์ Attention์ ๋ชจ๋ ๊ณ ๋ คํ์ฌ ์ฐ์ํ representation์ ์ป์ ์ ์์ต๋๋ค.
Proposed method
- ๋จผ์ Node-level attention ๋ Meta-path ๋ณ๋ก ๊ณ์ฐ๋ฉ๋๋ค. ๋ฐ๋ผ์ ์ฐ์ ์ ์ผ๋ก ๋์ผํ Metapath ์ ์ํ๋ Node type(๋ฐฐ์ฐ, ์ํ, ์๊ฐ)์ ์๋ก ๋ค๋ฅธ attribute set(eg. ๋ฐฐ์ฐ : {๋์ด, ์ฑ๋ณ} , ์ํ : {์ฅ๋ฅด, ์ฐ๋})์ Type-specific transformation matrix ๋ฅผ ํตํด ๋์ผํ ๊ณต๊ฐ์ Mapping ํฉ๋๋ค..
- Node pair์ attention์ ๊ตฌํ๊ธฐ ์ํด, ๊ฐ node i,j์ , ๋ฅผ concat(notation โ ||)ํ์ฌ single layer mlp์ softmax function์ ์ฌ์ฉํ attention score๋ฅผ ๊ณ์ฐํฉ๋๋ค. โ> bahdanau attention
- ๊ณ์ฐ๋ node pair attention score๋ฅผ ํ์ฉํ์ฌ metapath ์์ Node ์ Embedding ๋ฅผ ์ป์ต๋๋ค.
- ๋ชจ๋ Metapath ์์ ๊ฐ๊ฐ ์ป์ Node ์ Embedding ๋ก ๋ถํฐ metapath(semantic)-level attention ์ ๊ณ์ฐํ๊ธฐ ์ํด, ๋ค์ single layer mlp์ softmax function์ ์ฌ์ฉํฉ๋๋ค.
- node-level๊ณผ metapath(semantic)-level attention์ ๋ชจ๋ ๊ณ ๋ คํ Node ์ embedding ๋ ๋ค์๊ณผ ๊ฐ์ต๋๋ค. ํ์ต์ Cross-entropy loss๋ฅผ ์ฌ์ฉํ์ฌ ์งํ๋ฉ๋๋ค.
Experiment
๋ณธ ๋
ผ๋ฌธ์์ ์ ์ํ๋ Heterogeneous graph attention network(HAN)์ ACM, DBLP์ ๊ฐ์ citation network dataset๊ณผ IMDB dataset์์ Deepwalk, GCN, GAT์ ๋น๊ตํ์ฌ ์ฐ์ํ node classification score๋ฅผ ๊ธฐ๋กํ์์ต๋๋ค.