STAR-GCN: Stacked and Reconstructed Graph Convolutional Networks for Recommender Systems, IJCAI 19, Zhang et al

STAR-GCN: Stacked and Reconstructed Graph Convolutional Networks for Recommender Systems, IJCAI 19, Zhang et al

Goal
Inductive(and Transductive) User-Item rating(Matrix Completion) via GNN

Challenge

• Generalize model to predict unseen nodes embeddings for cold start problem
  • one-hot vector(embedding look-up table)를 쓰면 unseen node에 대한 generalize가 잘 안 되고, cold-start problem을 야기하는 문제 발생
• Label leakage issue
  • ex) user1과 item1 사이에 edge가 있을지 여부를 predict하는 task를 풀고 있는데, 애초에 aggregation 과정에서 neighbor의 표현을 토대로 update를 하기 때문에, label leakage가 발생하여 overfitting으로 이어짐

Solution

• Mask & Reconstruction strategy
• Sample & Remove strategy

Method

• STAR-GCN(STAcked and Reconstructed Graph ConvolutionalNetworks)
• Masked & Reconstruct
  • To generalize embedding for new nodes,
    1. masking some percentage of input nodes
    2. reconstructing the clean node embeddings
• Sample & Remove
  • rating pairs를 fixed size만큼 샘플링하고, 해당 edge를 graph에서 remove함
• Multi-Block graph Encoder-Decoder architecture
  • Block = Encoder + Decoder
    • Block Stacking : 각 블록 별로 별개의 parameter 사용
    • Recurrence : parameter sharing for across all blocks
  • Encoder
    • Generate node representations
    • Multi-link GCN aggregator (which is used in GC-MC)
  • Decoder
    • Recover input node embeddings
    • 2-layer fc layers
• Optimization
  • Jointly training for the rating prediction(main task) and reconstruction(auxiliary) loss

ETC.

• Inductive와 Transductive 둘 다 가능
  • 지난 번 논문도 그렇고, GNN쪽에서 inductive를 가능케 하기 위한 접근으로 embedding reconstruction을 취하는 경우가 많은 것 같네
• Task
  • Matrix Completion via GNN
    • Main task : link prediction
    • Auxiliary task : Embedding reconstruction
• GC-MC 훑어보기
• 깃허브로 블로그 이사하기!