Scaling Graph Neural Networks for Friend Recommendation: Multi-Hash User Embeddings and Temporal Neighbor Sampling

Maksim Utushkin, Andrei Ovsiannikov, Alexander D'yakonov|August 27, 2026arXiv

Key Takeaway

Multi-hash embeddings and timestamp-indexed neighbor sampling are practical techniques that make GNNs deployable on billion-scale social graphs without sacrificing ranking quality or inference speed.

Summary

This paper presents a production-scale GNN system for friend recommendations on massive social graphs (194M users, 28B edges). The key innovations are multi-hash embeddings that reduce ID storage by 98% while maintaining quality, and temporal neighbor sampling optimized to O(log(deg(v)) + k) complexity. The system achieved 16% more friend additions in live A/B testing.

scaling applications efficiency

Key Terms

graph-neural-network message-passing multi-hop temporal-neighbor-sampling embedding-table