Meta's RankGraph-2 slashes serving costs 83% while boosting CTR & CVR

Meta researchers present RankGraph-2, a lifecycle co-design framework that unifies graph construction, representation learning, and real-time serving for similarity-based retrieval at billion-node scale. Unlike prior work that tackles each stage in isolation, RankGraph-2 recognizes that serving constraints (avoiding expensive online KNN) drive the need for a co-learned cluster index baked into training objectives; training tolerates pre-computed neighborhoods, which in turn demands construction produce self-contained data with hour-level refresh. This tight coupling lets RankGraph-2 subsample hundreds of trillions of edges down to hundreds of billions using popularity bias correction, pre-compute multi-hop neighborhoods via personalized PageRank, and co-train a residual-quantization index that slashes serving computational cost by 83%.

Empirically, RankGraph-2 delivers 3.8× higher recall than a GAT+Deep Graph Infomax model on bipartite graphs and 2.1× higher than PyTorch-BigGraph on item retrieval. In production at Meta, it has powered over 20 retrieval launches across major surfaces, driving up to +0.96% click-through rate and +2.75% conversion rate. By demonstrating that a simple architecture can beat sophisticated models when lifecycle stages are co-designed, RankGraph-2 sets a new practical baseline for billion-scale graph retrieval.