A Node-Aware Dynamic Quantization Approach for Graph Collaborative Filtering

Lin Li, Chunyang Li, Yu Yin, Xiaohui Tao, Jianwei ZhangCIKM 2025 · DOI · OpenReview

TL;DR

This paper introduces GNAQ, a novel quantization approach for Graph Neural Networks (GNNs) in collaborative filtering that tackles the challenge of deploying these powerful models on resource-constrained devices. By dynamically adapting quantization parameters to individual nodes based on their graph structure, GNAQ significantly reduces model size and computational cost while maintaining high recommendation accuracy, outperforming existing methods.

What Problem Does This Paper Address?

Collaborative filtering (CF) models, especially those leveraging Graph Neural Networks (GNNs), have become indispensable for personalized recommendations. However, their success comes at a cost: large embedding parameter requirements and high computational demands. This makes deploying them on edge devices with limited resources a significant hurdle. Existing quantization techniques, which aim to compress models by using lower-precision numbers, often overlook the inherent graph structure of GNNs. Directly quantizing node embeddings without considering their relationships can lead to error accumulation during message passing, degrading recommendation quality. Furthermore, many current GNN quantization methods treat all nodes uniformly, failing to account for the diverse roles and connectivity patterns within a graph, leading to suboptimal performance.

Methods

The paper proposes Graph based Node-Aware Dynamic Quantization (GNAQ), a novel quantization-aware training approach for GNN-based collaborative filtering. GNAQ introduces three key innovations:

  • Node-Aware Parameter Sharing Quantization: This technique adapts quantization parameters to the local structure of each node. It leverages the idea of parameter sharing, where an allocation matrix is learned to associate node encodings with meta-embeddings, allowing for dynamic determination of quantization step sizes through backpropagation.
  • Dynamic Step Size Quantization: GNAQ dynamically refines quantization intervals based on individual node embedding distributions. It initializes quantization intervals using the maximum and minimum values of node embeddings and then iteratively updates them through message passing in GNN layers. This allows the quantization step size to adapt to the specific data distribution of each node.
  • Graph Interaction-Aware Quantization Updates: To mitigate information loss and improve accuracy, GNAQ incorporates graph structural information into the quantization process. Specifically, it uses the first-order aggregated information of neighboring nodes to update the node quantization encoding, providing a form of regularization and preventing issues like "Z-shaped" oscillation during parameter updates.

Additionally, GNAQ employs a graph relation-aware gradient estimation to replace traditional straight-through estimators, ensuring more accurate gradient propagation during training. The loss function also includes a ranking loss (combining BPR and LambdaLoss) to specifically enhance the model's ability to predict and rank relevant items.

Datasets

The proposed GNAQ method was evaluated on four widely used public datasets:

  • MovieLens
  • Gowalla
  • Yelp2020
  • Amazon-Book

The paper notes that Yelp2020 and Amazon-Book can be found in the official code repositories of LightGCN and LEGCF, respectively.

Key Insights and Ideas

  • Structure Matters for Quantization: The core insight is that simply applying generic quantization to GNNs ignores their graph structure, leading to performance degradation. GNAQ's success hinges on explicitly incorporating node relationships into the quantization process.
  • Dynamic Adaptation is Crucial: Unlike static quantization methods, GNAQ's ability to dynamically adjust quantization parameters (step sizes and intervals) for each node based on its local graph context is key to its effectiveness. This allows for a more nuanced and accurate representation of node embeddings.
  • Parameter Sharing as a Quantization Tool: The paper cleverly repurposes parameter sharing, a technique for reducing model size, to enable dynamic quantization. By learning an allocation matrix, GNAQ can infer quantization parameters without explicit, per-node parameter storage.
  • Gradient Estimation for Quantized GNNs: The use of graph relation-aware gradient estimation is vital for training quantized GNNs effectively, overcoming the non-differentiable nature of quantization operations.

Key Results

GNAQ demonstrates significant improvements over state-of-the-art quantization methods:

  • Accuracy: On average, GNAQ achieved a 27.8% improvement in Recall@10 and a 17.6% improvement in NDCG@10 under 2-bit quantization across experiments.
  • Efficiency: GNAQ can maintain the performance of full-precision models while reducing their model sizes by 8 to 12 times. The training time is also reported to be twice as fast compared to quantization baseline methods.
  • Ablation Studies: Experiments showed that removing dynamic quantization steps (DQS) or relational aggregation-based updates (RAU) led to substantial performance degradation, highlighting the importance of these components. The ranking-enhanced loss specifically improved ranking accuracy without compromising recall.

Limitations and Future Work

  • Limitations: The paper acknowledges that while GNAQ's parameter size is higher than some lightweight models like LEGCF, it achieves superior performance. The authors also note that on the Gowalla dataset, GNAQ had a slight NDCG@20 deficit compared to BiGeaR at 64 dimensions, potentially due to insufficient neighbor information affecting the Relational Aggregation-Based Update (RAU).
  • Future Work: The authors plan to apply GNAQ to actual edge deployment scenarios to achieve efficient inference on resource-constrained devices while maintaining high recommendation performance.

Summary updated

BibTeX
@inproceedings{li2025node,
  author = {Lin Li and Chunyang Li and Yu Yin and Xiaohui Tao and Jianwei Zhang},
  title = {{A Node-Aware Dynamic Quantization Approach for Graph Collaborative Filtering}},
  booktitle = {ACM International Conference on Information and Knowledge Management},
  year = {2025},
  doi = {10.1145/3746252.3761244},
}
Figure