This paper presents PinSage, a scalable graph convolutional network (GCN) designed for web-scale recommender systems. The authors developed PinSage at Pinterest, a platform with billions of items and users, to generate high-quality recommendations. PinSage combines efficient random walks and graph convolutions to generate embeddings that incorporate both graph structure and node features. It uses a novel method based on efficient random walks to structure convolutions and a novel training strategy that relies on harder-and-harder training examples to improve robustness and convergence. PinSage is deployed on a graph with 3 billion nodes and 18 billion edges, and trained on 7.5 billion examples. According to offline metrics, user studies, and A/B tests, PinSage generates higher-quality recommendations than comparable deep learning and graph-based alternatives. This is the largest application of deep graph embeddings to date and paves the way for a new generation of web-scale recommender systems based on graph convolutional architectures. The key innovations of PinSage include on-the-fly convolutions, producer-consumer minibatch construction, and efficient MapReduce inference. These techniques improve the scalability of GCNs and allow for efficient training on large graphs. Additionally, the paper introduces new training techniques and algorithmic innovations that improve the quality of the representations learned by PinSage, leading to significant performance gains in downstream recommender system tasks. The authors also evaluate PinSage against several baselines, including visual embeddings, annotation embeddings, and graph-based methods. The results show that PinSage outperforms these baselines in terms of hit rate and MRR. User studies further confirm that PinSage generates more relevant recommendations than other methods. Production A/B tests show that PinSage improves user engagement on the homefeed recommendation task. The paper concludes that PinSage demonstrates the impact of graph convolutional methods in a production recommender system and that it can be further extended to tackle other graph representation learning problems at large scale, including knowledge graph reasoning and graph clustering.This paper presents PinSage, a scalable graph convolutional network (GCN) designed for web-scale recommender systems. The authors developed PinSage at Pinterest, a platform with billions of items and users, to generate high-quality recommendations. PinSage combines efficient random walks and graph convolutions to generate embeddings that incorporate both graph structure and node features. It uses a novel method based on efficient random walks to structure convolutions and a novel training strategy that relies on harder-and-harder training examples to improve robustness and convergence. PinSage is deployed on a graph with 3 billion nodes and 18 billion edges, and trained on 7.5 billion examples. According to offline metrics, user studies, and A/B tests, PinSage generates higher-quality recommendations than comparable deep learning and graph-based alternatives. This is the largest application of deep graph embeddings to date and paves the way for a new generation of web-scale recommender systems based on graph convolutional architectures. The key innovations of PinSage include on-the-fly convolutions, producer-consumer minibatch construction, and efficient MapReduce inference. These techniques improve the scalability of GCNs and allow for efficient training on large graphs. Additionally, the paper introduces new training techniques and algorithmic innovations that improve the quality of the representations learned by PinSage, leading to significant performance gains in downstream recommender system tasks. The authors also evaluate PinSage against several baselines, including visual embeddings, annotation embeddings, and graph-based methods. The results show that PinSage outperforms these baselines in terms of hit rate and MRR. User studies further confirm that PinSage generates more relevant recommendations than other methods. Production A/B tests show that PinSage improves user engagement on the homefeed recommendation task. The paper concludes that PinSage demonstrates the impact of graph convolutional methods in a production recommender system and that it can be further extended to tackle other graph representation learning problems at large scale, including knowledge graph reasoning and graph clustering.