Graph Convolutional Neural Networks for Web-Scale Recommender Systems

This paper describes a large-scale deep recommendation engine developed and deployed at Pinterest, tackling the challenge of making graph convolutional networks practical at web scale. The core method is a data-efficient Graph Convolutional Network that combines efficient random walks with graph convolutions to generate embeddings of items that incorporate both the graph structure and node feature information. The authors introduce a novel random-walk-based scheme to structure the convolutions, a training strategy that relies on progressively harder examples to improve robustness and convergence, and an efficient MapReduce inference algorithm for generating embeddings.

These techniques let the system train on and embed graphs four orders of magnitude larger than typical GCN implementations, operating on Pinterest's underlying graph of 3 billion nodes and 17 billion edges. Offline metrics, user studies, and A/B tests all indicate the approach produces higher-quality recommendations than comparable deep-learning-based systems. The authors describe it as by far the largest application of deep graph embeddings to date, one that paves the way for a new generation of web-scale recommenders based on graph convolutional architectures.