Meta rolled out RankGraph-2, a system that designs graph construction, representation learning and serving as a single pipeline. The framework trims a billion‑node graph from hundreds of trillions of edges down to a few hundred billion by subsampling with popularity‑bias correction and pre‑computes multi‑hop neighborhoods using personalized PageRank. A residual‑quantization cluster index is co‑trained with the model, removing the need for costly online K‑nearest‑neighbour searches and cutting serving compute by 83%.
The changes matter because existing large‑scale recommendation pipelines treat each stage in isolation. By forcing the stages to accommodate one another, RankGraph-2 achieves 3.8 × higher recall than a GAT + Deep Graph Infomax baseline on a bipartite graph and 2.1 × higher than PyTorch‑BigGraph on item retrieval. In production it adds up to 0.96 % in click‑through rate and 2.75 % in conversion rate, enough to justify 20+ launches across Meta’s major surfaces.
The approach echoes earlier attempts at end‑to‑end graph pipelines, but those typically kept a heavy online graph service that limited refresh speed. RankGraph-2’s hour‑level refresh and self‑contained neighborhood data sidestep that bottleneck, putting it ahead of static‑graph methods that struggle with item coverage. Competitors still reliant on full‑graph KNN will face higher latency and infrastructure costs.
In short, Meta’s co‑design shows that shaving off needless edges and merging index training into the model can deliver measurable business lift without a wholesale architecture overhaul.