Emerging
Jun 18, 20261
66%
Meta Introduces RankGraph-2 Framework for Billion-Node Graph-Based Recommendation Retrieval
Meta researchers have introduced RankGraph-2, a framework that co-designs graph construction, representation learning, and real-time serving for billion-node scale recommendation systems. The system achieves significant performance improvements (3.8x higher recall than comparable models) and has powered over 20 retrieval launches at Meta, delivering measurable gains in click-through rates and conversion rates.
Quick Facts
Who
Meta researchers
What
Developed RankGraph-2 framework for billion-node graph-based retrieval
When
Submitted June 16, 2026
Where
Meta
- Developed RankGraph-2 framework for billion-node graph-based retrieval
- Co-designed graph construction, representation learning, and real-time serving stages
- Reduced hundreds of trillions of edges to hundreds of billions via subsampling
- Pre-computes multi-hop neighborhoods using personalized PageRank
- Co-learns residual-quantization cluster index
Meta researchers have unveiled RankGraph-2, a new framework designed to handle graph-based retrieval at billion-node scale for recommendation systems. The framework addresses a long-standing challenge in machine learning by co-designing three tightly coupled problems—graph construction, representation learning, and real-time serving—that typically have been tackled in isolation. The research was submitted to arXiv on June 16, 2026, in the Computer Science > Information Retrieval category.
The RankGraph-2 framework is designed to optimize similarity-based retrieval patterns including user-to-user-to-item (U2U2I) and user-to-item-to-item (U2I2I) retrieval. A core innovation is the use of a co-learned cluster index to eliminate expensive online k-nearest neighbor (KNN) operations during serving, which pushes index co-training directly into the training objective. The system leverages pre-computed neighborhoods rather than requiring online graph infrastructure, allowing the construction stage to produce self-contained data. The framework also supports hour-level refresh cycles to maintain current item coverage.
To manage the scale challenge, RankGraph-2 employs several key techniques. It reduces hundreds of trillions of edges to hundreds of billions through subsampling with popularity bias correction, pre-computes multi-hop neighborhoods using personalized PageRank, and co-learns a residual-quantization cluster index. These optimizations achieve an 83% reduction in serving computational cost. In performance benchmarks, RankGraph-2 delivers 3.8 times higher recall than a Graph Attention Network (GAT) plus Deep Graph Infomax model on bipartite graphs, and 2.1 times higher recall compared to PyTorch-BigGraph on item retrieval tasks.
The framework has demonstrated real-world impact at Meta. Deployment has resulted in measurable improvements to key business metrics, including up to 0.96% improvement in click-through rate (CTR) and 2.75% improvement in conversion rate (CVR). The system has powered more than 20 retrieval launches across major Meta surfaces, indicating broad adoption and validation within the company's recommendation infrastructure.
Why This Matters
RankGraph-2 represents a major advancement in recommendation system architecture that directly impacts user experience at scale. For practitioners, the framework demonstrates how to unify typically isolated optimization stages (graph construction, learning, serving) into a coherent system that achieves 83% serving cost reduction while delivering measurable business impact—up to 0.96% CTR improvement and 2.75% CVR improvement. This is actionable for any organization deploying large-scale recommendation systems, as it provides a production-proven methodology for handling billion-node graphs efficiently.
Timeline & Sources
Jun 16, 2026
WireRankGraph-2 research paper submitted to arXiv
Jun 18, 2026
WireRankGraph-2 paper published and announced on arXiv