🏷️

PinSage: Graph Convolutional Neural Networks for Web-Scale Recommender Systems

Key Points

GCN-based PinSage combines efficient random walks and graph convolutions to generate embeddings of user-item pairs.
Embeddings incorporate graph structure as well as node feature information.
Trained using a pairwise ranking loss function and able to handle sparsity by using negative sampling.
- Negative sampling is used to generate negative examples for the pairwise ranking loss function.
- These examples are generated by randomly sampling items that are not connected to the user in the graph structure of the data.
Pairwise ranking loss function:
- Two parts — positive: loss for the correct ordering of the items.
- Two parts — negative: loss for the incorrect ordering of the items.
- Total loss: sum of both losses.
Trained on 3 billion nodes representing pins and boards and 18 billion edges — according to A/B tests, PinSage generates higher quality recommendations than other deep learning alternatives.

PinSage architecture overview: pins and boards as nodes in a bipartite graph with convolution-based aggregation.

Importance Pooling

Feature information is aggregated from local neighborhoods in the graph — but PinSage introduces a method to weigh the importance of node features, based upon random-walk similarity measures.

PinSage importance pooling algorithm showing how random walk similarities weight neighborhood aggregation.

Loss

PinSage loss function: pairwise max-margin ranking loss combining positive and negative edge scores.