A new GNN ranking system addresses the challenges of deploying friend recommendation models on social graphs containing hundreds of millions of users and tens of billions of connections. The system, detailed in a paper on arXiv, integrates multi-hash ID embeddings and temporal neighbor sampling to handle the scale and dynamic nature of social networks. These design choices aim to optimize memory usage and improve the relevance of recommendations by considering the evolving social context.
Friend recommendation systems inherently rely on graph structures, where the relevance of a potential connection is often determined by multi-hop social context rather than isolated user attributes. Graph Neural Networks are well-suited for such tasks as they can capture complex relationships and propagate information across a graph to generate node embeddings. However, applying traditional GNNs to production-scale social graphs presents significant modeling and systems challenges, particularly concerning memory consumption for user identification and the need to reflect changes over time.
One key innovation in this system is the use of multi-hash ID embeddings. In large-scale recommender systems, user and item IDs are high-cardinality features, meaning there are many unique IDs. Storing full embedding tables for these IDs can require over 200 GB of memory for graphs of this size. Multi-hash embeddings address this by mapping multiple IDs to the same bucket, thereby reducing memory requirements. This approach allows the system to manage large numbers of unique identifiers without exceeding memory capacity, a common issue for industrial GNN deployments that often either omit trainable IDs or accept the memory cost of full embedding tables. Previous research has explored hashing-based approaches to reduce memory for embedding layers in GNNs, with some methods achieving substantial memory reductions while maintaining or improving accuracy.
The system also incorporates temporal neighbor sampling. Social graphs are dynamic, with new connections forming and old ones changing constantly. Traditional GNNs often struggle to effectively model these temporal evolutions. Temporal neighbor sampling allows the system to adaptively select relevant neighbors for a node at any given time, capturing the dynamic nature of user interactions. This is crucial for friend recommendation, where the recency and sequence of interactions can significantly influence the likelihood of a new connection. Existing temporal GNNs (TGNNs) have shown improved performance by integrating time information, but often require specialized models and training frameworks. The proposed system aims to integrate temporal awareness in a scalable manner, reflecting how user interests and social influences evolve over time.
The development of this system builds on ongoing research into scaling GNNs for various applications. For instance, efforts have been made to scale GNN training for link prediction tasks by developing methods that reduce computational cost and memory footprint through strategies like self-sufficient partitions and edge mini-batch training. Other work has focused on learning discrete representations using GNNs for efficient retrieval in large-scale recommendation systems, demonstrating comparable performance to continuous embedding counterparts but with faster inference. The integration of multi-hash embeddings and temporal sampling in this new system represents a specific approach to these broader challenges within the context of friend recommendation.
The researchers indicate that the system provides an end-to-end GNN ranking solution, which is critical for real-world deployment. The combination of memory-efficient embeddings and dynamic neighbor sampling allows the model to process the vast and ever-changing data of a large social network, offering a pathway for more accurate and timely friend suggestions.