6 Jun 2018 | Rex Ying*,†, Ruining He*, Kaifeng Chen*†, Pong Eksombatchai*, William L. Hamilton†, Jure Leskovec*†
This paper introduces PinSage, a scalable Graph Convolutional Network (GCN) designed for web-scale recommender systems. PinSage combines efficient random walks and graph convolutions to generate node embeddings that incorporate both graph structure and node feature information. The authors address the challenge of scaling GCNs to large graphs with billions of nodes and edges by developing a novel training strategy and efficient inference methods. PinSage is deployed at Pinterest, where it outperforms other deep learning and graph-based alternatives in various recommendation tasks, including item-item recommendations and homefeed recommendations. Extensive offline metrics, user studies, and A/B tests demonstrate the effectiveness of PinSage, showing improvements of up to 40% in hit rates and 60% in Mean Reciprocal Rank (MRR). The paper also discusses the scalability and efficiency of PinSage, highlighting its ability to handle large datasets and its potential for future applications in graph representation learning.This paper introduces PinSage, a scalable Graph Convolutional Network (GCN) designed for web-scale recommender systems. PinSage combines efficient random walks and graph convolutions to generate node embeddings that incorporate both graph structure and node feature information. The authors address the challenge of scaling GCNs to large graphs with billions of nodes and edges by developing a novel training strategy and efficient inference methods. PinSage is deployed at Pinterest, where it outperforms other deep learning and graph-based alternatives in various recommendation tasks, including item-item recommendations and homefeed recommendations. Extensive offline metrics, user studies, and A/B tests demonstrate the effectiveness of PinSage, showing improvements of up to 40% in hit rates and 60% in Mean Reciprocal Rank (MRR). The paper also discusses the scalability and efficiency of PinSage, highlighting its ability to handle large datasets and its potential for future applications in graph representation learning.