PinSage: A new graph convolutional neural network for web-scale recommender systems

Abstraction: Web-scale random-walk GCN for billion-node Pinterest recommender system

Key points:

• PinSage is a random-walk Graph Convolutional Network operating on a bipartite Pin-board graph with 3 billion nodes and 18 billion edges — 10,000× larger than typical GCN applications
• Key innovation: "on-the-fly" localized convolutions via dynamically constructed computation graphs instead of full graph Laplacian multiplication
• Neighborhoods defined by random-walk visit counts (importance pooling) rather than k-hop neighborhood, yielding a 46% offline metric gain over k-hop baselines
• MapReduce-based inference generates embeddings for billions of nodes in a few hours on a cluster of a few hundred instances
• Outperformed top content-based baseline by 40% absolute (150% relative) in Recall and 22% absolute (60% relative) in MRR
• A/B tests showed ~30% relative improvement in user engagement rates in Home Feed and Related Pin Ads

Connections: Pinterest · Graph Neural Networks · Recommender Systems · Graph Embeddings

Source: https://medium.com/pinterest-engineering/pinsage-a-new-graph-convolutional-neural-network-for-web-scale-recommender-systems-88795a107f48