GraphHash: Graph Clustering Enables Parameter Efficiency in Recommender Systems

TL;DR

GraphHash addresses embedding-table memory growth in deep recommender systems by performing modularity-based bipartite graph clustering on the user–item interaction graph to generate bucket assignments for users and items. The method links modularity maximization to a smoothing effect akin to message-passing, offering a scalable preprocessing alternative via the Louvain algorithm and a plug-and-play hashing mechanism. Empirical results show GraphHash substantially improves retrieval metrics (Recall and NDCG) with fewer parameters, while a DoubleGraphHash variant further enhances CTR performance, indicating practical gains in production settings. Overall, GraphHash provides a robust, graph-informed, parameter-efficient approach to embedding reduction that aligns with industrial needs for efficiency and scalability.

Abstract

Deep recommender systems rely heavily on large embedding tables to handle high-cardinality categorical features such as user/item identifiers, and face significant memory constraints at scale. To tackle this challenge, hashing techniques are often employed to map multiple entities to the same embedding and thus reduce the size of the embedding tables. Concurrently, graph-based collaborative signals have emerged as powerful tools in recommender systems, yet their potential for optimizing embedding table reduction remains unexplored. This paper introduces GraphHash, the first graph-based approach that leverages modularity-based bipartite graph clustering on user-item interaction graphs to reduce embedding table sizes. We demonstrate that the modularity objective has a theoretical connection to message-passing, which provides a foundation for our method. By employing fast clustering algorithms, GraphHash serves as a computationally efficient proxy for message-passing during preprocessing and a plug-and-play graph-based alternative to traditional ID hashing. Extensive experiments show that GraphHash substantially outperforms diverse hashing baselines on both retrieval and click-through-rate prediction tasks. In particular, GraphHash achieves on average a 101.52% improvement in recall when reducing the embedding table size by more than 75%, highlighting the value of graph-based collaborative information for model reduction. Our code is available at https://github.com/snap-research/GraphHash.

Figures (7)

• Figure 1: Overview of GraphHash. By employing fast graph clustering algorithms, GraphHash serves as a computationally efficient proxy for message-passing during preprocessing and a plug-and-play graph-based alternative to traditional ID hashing, working seamlessly with any architectural backbone that utilizes embedding tables.
• Figure 2: Performance breakdown of the retrieval task by test user frequency in the training data. Frequency information tends to benefit power users, regardless of the backbone model. In contrast, GraphHash achieves balanced performance across all user groups, closely mirroring the trend of the full model.
• Figure 3: Performance breakdown of the CTR task by user frequency in training data. All methods tend to perform better for clicks generated by power users, and DoubleGraphHash, which obtains the best overall performance, also works better for clicks generated by power users than for those generated by tail users.
• Figure 4: Impact of the uniformity term $\gamma$ in DirectAU on model performance. While the full model and GraphHash are robust to changes in $\gamma$, double frequency hashing shows a sweet spot, suggesting GraphHash enhances robustness to $\gamma$ in hashing methods.
• Figure 5: The impact of LightGCN's depth on the performance of different hashing methods. GraphHash consistently outperforms random hashing. In particular, GraphHash without any additional message-passing layers, performs roughly equal to random hashing with one or two message-passing layers, where the performance in the latter model can be sorely attributed to pure message-passing.

Theorems & Definitions (1)

• proposition 1