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

Based on the provided search context, there is no record of a specific paper titled "Scaling Graph Neural Networks for Friend Recommendation: Multi-Hash User Embeddings and Temporal Neighbor Sampling." The context does, however, detail several other scalable GNN systems for friend recommendation that utilize temporal neighbor sampling and address similar production challenges:

• GraFRank (Sankar et al., WWW 2021): A multi-faceted friend ranking architecture that employs modality-specific neighbor aggregators and temporal neighborhood sampling to learn expressive user representations. Tested on Snapchat’s multi-million user datasets, it outperformed baselines by 30% MRR in candidate retrieval and 20% MRR in ranking, with significant gains for less-active and low-degree users.
• GiGL (Snap Research, 2025): An infrastructure library powering GNN workflows at Snapchat for friend recommendation, content recommendation, and ads. It supports graphs with ~900M nodes and ~16.8B edges, utilizing both tabularized and real-time subgraph sampling. GiGL has enabled over 35 production launches and incorporates supervised link prediction setups using newly established friendships as positive labels.
• SSNet (Huang et al., 2022): A self-rescaling network that bridges neighborhood aggregator normalization with end-to-end representation learning. Applied to Xbox friend recommendation, it improved online A/B test metrics by 24% and addressed scale distortion issues in GNN representations.
• LTGNN (2024): A Linear-Time Graph Neural Network for scalable recommendations that uses a single propagation layer and variance-reduced neighbor sampling to achieve linear complexity while maintaining expressive power for capturing long-range collaborative signals.

The specific technique of "Multi-Hash User Embeddings" is not mentioned in the provided context.

The paper addresses the practical difficulty of applying graph neural networks to friend recommendation on industrial-scale social graphs, where candidate users, interactions, and friendship edges must be modeled across hundreds of millions of nodes. Its central goal is to make multi-hop GNN ranking feasible in production rather than in a simplified batch setting. The proposed system combines multi-hash user embeddings with temporal neighbor sampling, allowing the model to represent users compactly and to expand neighborhoods in a way that emphasizes recent, relationship-relevant structure instead of treating the entire static graph as equally useful.

The key contribution is a systems-aware formulation of GNN-based friend ranking. Multi-hash embeddings provide a scalable representation strategy for large ID spaces, supporting efficient similarity and ranking over massive candidate sets without relying on prohibitively large dense feature stores. Temporal neighbor sampling addresses the nonstationarity of social graphs: friendship propensity is strongly influenced by recent activity, evolving communities, and changing interaction patterns, so the model prioritizes temporally meaningful neighbors while controlling the computational cost of multi-hop expansion. Together, these components form an end-to-end approach that balances graph fidelity, memory usage, and serving latency.

This work matters because friend recommendation is a high-impact link-prediction problem, yet many GNN methods are difficult to deploy when the underlying graph is enormous and constantly changing. By focusing on production constraints—neighborhood explosion, temporal drift, embedding storage, and ranking efficiency—the paper offers a practical template for scaling GNNs to other large-graph recommendation settings, such as creator, group, content, or job matching. Its main insight is that scalable graph learning at this scale requires co-designing the model, sampling strategy, and embedding representation rather than treating graph neural networks as a purely algorithmic layer.