RankGraph-2: Scalable Industrial Graph Retrieval¶

Core-mechanism reproduction. Popularity correction, offline PPR, and cluster indexing execute.

Paper information¶

Field	Value
Paper	arXiv 2606.18379
Affiliations	Meta
First public release	2026-06-16 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`rankgraph2`
Local code	`src/auto_research/reproductions/rankgraph2/`

Original paper¶

Background and contribution¶

Online multi-layer GNNs are costly and can overemphasize popular edges. Correct edge popularity, precompute multihop PPR, and compress serving with a jointly learned residual cluster index.

flowchart LR
  A[User-item graph] --> B[Corrected edges] --> C[Offline multihop PPR] --> D[Residual cluster index] --> E[Low-cost retrieval]

Key figure from the paper¶

Original Figure 2: graph retrieval and serving. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ \pi_i=\alpha e_i+(1-\alpha)\pi_i\tilde P,\qquad \tilde P_{uv}\propto P_{uv}/\sqrt{1+\operatorname{pop}(v)}. \]

Paper results¶

The paper reports 83% less compute, 3.8×/2.1× recall, and production CTR +0.96%, CVR +2.75%, across more than 20 launches.

Local reproduction¶

On a MovieLens item graph, debiased transitions, multihop PPR, and PPR clustering run. Against one-hop retrieval, debiased PPR plus two-stage residual cluster indexing gives NDCG@10 0.03540→0.07421 (+109.65%), in metrics/movielens-100k-seed42.json.

auto-research reproduce --paper rankgraph2 --dataset-dir data --seed 42

Reproduction limits¶

MovieLens evidence is distinct from the paper's online results.