Source linked

Meta's RankGraph-2 Cuts Graph Serving Cost 83% at Billion-Node Scale

By co-designing graph construction, training, and serving as one pipeline, RankGraph-2 achieves 3.8x higher recall and up to +2.75% CVR across 20+ Meta surfaces.

metarankgraph 2graph learningrecommendation systemsbillion scale retrievalsemantic similarity

Meta deployed RankGraph-2 for billion-node graph learning in recommendation, and the numbers are concrete: serving computational cost drops 83%, recall jumps 3.8x over a GAT + Deep Graph Infomax baseline, and CTR lifts up to 0.96% with CVR up to 2.75%. Twenty-plus retrieval launches across major surfaces are already running on it.

Three Problems, One Co-Designed Pipeline

Most graph retrieval systems treat graph construction, representation learning, and serving as separate silos. RankGraph-2 throws that out. It treats all three stages as coupled constraints: serving demands a co-learned cluster index to avoid expensive online KNN, so that index is baked into the training objective. Training exploits the fact that similarity-based retrieval tolerates pre-computed neighborhoods, eliminating the need for online graph infrastructure. Construction must support hour-level refresh for item coverage, forcing subsampling and pre-computation.

How the Math Changes the Architecture

Hundreds of trillions of raw edges get subsampled to hundreds of billions using popularity bias correction. Multi-hop neighborhoods are pre-computed via personalized PageRank. A co-learned residual-quantization (RQ) cluster index replaces brute-force KNN, delivering that 83% serving cost reduction. The index isn't a bolt-on; it's co-trained with the embedding model so that cluster assignment aligns with the retrieval objective.

Measured Impact Across the Stack

On a bipartite graph, RankGraph-2 achieves 3.8x higher recall than a GAT + Deep Graph Infomax model. For item retrieval, it beats PyTorch-BigGraph by 2.1x. Real-world metrics are just as sharp: up to +0.96% CTR and +2.75% CVR. The framework has powered 20+ launches across Meta's major surfaces, which means it's not a toy paper experiment.

What This Enables Next

RankGraph-2 proves that lifecycle co-design isn't just an academic nice-to-have; it delivers production wins at billions of nodes. The pre-computed neighborhood and co-trained index pattern opens the door for graph-based retrieval at scales that previously required expensive online infrastructure. Expect more systems to copy this co-design pattern, because nobody can afford to leave an 83% serving cost savings on the table.


Source: RankGraph-2: Lifecycle Co-Design for Billion-Node Graph Learning in Recommendation
Domain: arxiv.org

Read original source ->

External source stays available while the OJO article and comment thread stay local.

Comments load interactively on the live page.