# Graph Convolutional Neural Networks for Web-Scale Recommender Systems

**Authors:** Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, Jure Leskovec  
**arXiv:** [1806.01973](https://arxiv.org/abs/1806.01973)  
**Date:** 2018-06-06  
**Summary page:** https://www.summarizepaper.com/en/paper/1806.01973v1/

## Key points

- Recent advancements in deep neural networks for graph-structured data have shown promising results in recommender system benchmarks.
- PinSage is a large-scale deep recommendation engine developed and deployed at Pinterest.
- PinSage utilizes a data-efficient Graph Convolutional Network (GCN) algorithm that combines efficient random walks and graph convolutions to generate embeddings of nodes.
- The embeddings incorporate both graph structure and node feature information, resulting in more accurate recommendations.
- PinSage introduces a novel method based on highly efficient random walks to structure the convolutions.
- PinSage employs a unique training strategy that relies on harder-and-harder training examples to improve the robustness and convergence of the model.
- To generate embeddings using a trained model, PinSage employs an efficient MapReduce model inference algorithm.
- PinSage is trained on an extensive dataset consisting of 7.5 billion examples on a graph with 3 billion nodes representing pins and boards, along with 18 billion edges.
- According to offline metrics, user studies, and A/B tests conducted by Pinterest, PinSage outperforms comparable deep learning and graph-based alternatives in terms of generating higher-quality recommendations.
- This research paves the way for future web-scale recommender systems based on graph convolutional architectures.

## Summary

Recent advancements in deep neural networks for graph-structured data have shown promising results in recommender system benchmarks. To address the challenge of scaling these methods to web-scale recommendation tasks with billions of items and hundreds of millions of users, the authors present PinSage—a large-scale deep recommendation engine developed and deployed at Pinterest. PinSage utilizes a data-efficient Graph Convolutional Network (GCN) algorithm that combines efficient random walks and graph convolutions to generate embeddings of nodes (i.e., items). These embeddings incorporate both graph structure and node feature information, resulting in more accurate recommendations. Compared to previous GCN approaches, PinSage introduces a novel method based on highly efficient random walks to structure the convolutions. Additionally, the authors design a unique training strategy that relies on harder-and-harder training examples to improve the robustness and convergence of the model. To generate embeddings using a trained model, PinSage employs an efficient MapReduce model inference algorithm. The authors deploy PinSage at Pinterest and train it on an extensive dataset consisting of 7.5 billion examples on a graph with 3 billion nodes representing pins and boards, along with 18 billion edges. According to offline metrics, user studies, and A/B tests conducted by Pinterest, PinSage outperforms comparable deep learning and graph-based alternatives in terms of generating higher-quality recommendations. This application of deep graph embeddings is currently the largest known implementation to date and paves the way for future web-scale recommender systems based on graph convolutional architectures. Overall, this research addresses the challenge of making deep neural networks practical and scalable for web-scale recommendation tasks by introducing PinSage—a large-scale deep recommendation engine that leverages efficient random walks and novel training strategies to generate high-quality recommendations based on graph-structured data.