RankGraph-2: A Unified Approach to Graph Learning for Recommendations

The challenge of graph-based retrieval at the scale of billions of nodes involves tackling three interrelated issues: graph construction, representation learning, and real-time serving. Traditional methods often treat these aspects separately, leading to inefficiencies. RankGraph-2, developed by Meta, presents an integrated framework that simultaneously addresses these lifecycle components for similarity-based retrieval, including user-to-user and user-to-item interactions. The framework's design ensures that the needs of one stage influence the others. For instance, the serving phase employs a co-learned cluster index to mitigate the high costs associated with online K-nearest neighbors (KNN), integrating this into the training process. Additionally, the training phase leverages the fact that similarity-based retrieval can function effectively with pre-computed neighborhoods, thus minimizing the need for extensive online graph infrastructure. The construction phase is designed to refresh item coverage on an hourly basis. By implementing subsampling techniques with popularity bias correction, RankGraph-2 reduces the number of edges from hundreds of trillions to hundreds of billions. It also pre-computes multi-hop neighborhoods using personalized PageRank and co-learns a residual-quantization cluster index, resulting in an 83% reduction in serving costs. This innovative lifecycle co-design achieves a recall rate 3.8 times higher than that of a GAT + Deep Graph Infomax model and 2.1 times higher than PyTorch-BigGraph, contributing to significant increases in click-through rates and conversion rates across various platforms.