VK Friend-GNN: Multi-hash embeddings and temporal neighbor sampling¶

Core multi-hash and temporal cutoff proxy under matched candidates.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	AI VK.
First public release	2026-08-27 (arXiv v1)
Author code	makut/VK-GNN.
Local adapter or method key	`friend-gnn`
Local reproduction code	`src/auto_research/reproductions/friend_gnn/`

Paper method and evidence¶

Use three independent hashes into a shared table, concatenate/project, and retrieve only historical neighbors via timestamp-sorted CSR and binary search. GATv2 and separate query/candidate heads score links.

flowchart LR
  A[Hash IDs into shared table] --> B[Select historical neighbors]
  B[Select historical neighbors] --> C[Graph scoring heads]

Original paper figure¶

Figure shows hashing and temporal neighbors. From the original paper; copyright remains with the authors.

Core equation¶

\[ e_u=T_{h_1(u)}\|T_{h_2(u)}\|T_{h_3(u)},\qquad p=\operatorname{lower\_bound}(ts_u,\tau-\Delta). \]

Reported results¶

Paper: 194M users, 28B edges, ID table compression over 98%. Production A/B against a strong baseline: friend additions +16%, unique friend-adding users +11.5%.

Local reproduction¶

auto-research reproduce --paper friend-gnn --dataset-dir data --seed 42

Same MovieLens-1M full catalog, popularity baseline vs temporal graph: Hit@10 0.0656→0.2281 (+247.62%), NDCG@10 0.0387→0.1100. Metrics. The item-transition graph is not a friendship graph; no VK logs, distributed GATv2 or production lift.