Eco-Aware Graph Neural Networks for Sustainable Recommendations

Recommender systems play a crucial role in alleviating information overload by providing personalized recommendations tailored to users’ preferences and interests. Recently, Graph Neural Networks (GNNs) have emerged as a promising approach for recommender systems, leveraging their ability to effectively capture complex relationships and dependencies between users and items by representing them as nodes in a graph structure. In this study, we investigate the environmental impact of GNN-based recommender systems, an aspect that has been largely overlooked in the literature. Specifically, we conduct a comprehensive analysis of the carbon emissions associated with training and deploying GNN models for recommendation tasks. We evaluate the energy consumption and carbon footprint of different GNN architectures and configurations, considering factors such as model complexity, training duration, hardware specifications and embedding size. By addressing the environmental impact of resource-intensive algorithms in recommender systems, this study contributes to the ongoing efforts towards sustainable and responsible artificial intelligence, promoting the development of eco-friendly recommendation technologies that balance performance and environmental considerations. Code is available at: https://github.com/antoniopurificato/gnn_recommendation_and_environment.

💡 Research Summary

The paper “Eco‑Aware Graph Neural Networks for Sustainable Recommendations” addresses a largely overlooked aspect of modern recommender systems: their environmental impact. While Graph Neural Networks (GNNs) have become a dominant technique for capturing high‑order relationships between users and items, the energy consumption and resulting carbon emissions of training and deploying such models have received little systematic study.

The authors first situate their work within the broader AI sustainability literature, noting that most prior carbon‑footprint analyses focus on natural‑language processing, computer vision, or information‑retrieval tasks. Only a handful of studies have examined recommender systems, and none have combined GNN‑based models with a detailed investigation of how architectural choices—especially embedding dimensionality—affect emissions.

Methodologically, the study relies on the open‑source CodeCarbon library to monitor power draw from both CPU and GPU every 30 seconds during training. Power usage is converted to CO₂‑equivalent (CO₂‑eq) using a fixed emission factor, providing a standardized metric that can be compared across studies. Four representative GNN‑based collaborative‑filtering models are evaluated: NGCF, LightGCN, SimGCL, and LightGCL. Each model is trained with four embedding sizes (32, 64, 128, 256) to explicitly assess the trade‑off between representation capacity and environmental cost.

All experiments are conducted on a single NVIDIA RTX A6000 GPU (10,752 CUDA cores, 48 GB RAM) using the RecBole framework. Training runs for 400 epochs with batch sizes of 2048 (training) and 4096 (validation), optimized by Adam (learning rate 0.001). No early stopping is applied, ensuring that the full training trajectory is captured for each configuration. Three public datasets are used: MovieLens‑1M, Amazon Beauty, and DianPing, each pre‑processed into an implicit feedback setting and split chronologically (latest interaction for test, second‑latest for validation).

Performance is measured with standard ranking metrics—Precision@K, Recall@K, NDCG@K, and HIT@K (K = 10, 100)—while environmental impact is reported as total CO₂‑eq (kg) per model run. Results show that LightGCN consistently achieves the highest recommendation quality across all datasets, confirming prior claims about its effectiveness despite its “lightweight” design. NGCF, in contrast, records the lowest carbon emissions on two of the three datasets, likely because its architecture involves fewer propagation steps and no additional augmentation. SimGCL and LightGCL, which incorporate random noise or SVD‑based graph augmentations, exhibit higher computational overhead and consequently larger emissions.

A clear monotonic relationship emerges between embedding size and CO₂‑eq: larger embeddings increase memory bandwidth and FLOPs, leading to higher power draw. For example, LightGCN’s emissions rise from roughly 4 kg at 32 dimensions to over 12 kg at 256 dimensions on the Beauty dataset. This pattern holds for all models, underscoring embedding dimensionality as a key lever for balancing accuracy against sustainability.

The authors discuss the practical implications of these findings. In scenarios where energy budgets are constrained—such as edge deployments or large‑scale industrial recommender pipelines—choosing a model like LightGCN with moderate embedding size may provide the best performance‑efficiency trade‑off. They also acknowledge several limitations: the use of a single GPU limits generalizability to multi‑GPU or distributed training environments; the emission factor is fixed and does not reflect regional variations in electricity generation; and only three datasets are examined, which may not capture the full diversity of real‑world recommendation tasks.

Future work is suggested to extend the analysis to other hardware platforms (TPUs, ASICs), incorporate dynamic power‑pricing or renewable‑energy aware scheduling, and explore model‑compression techniques (pruning, quantization) as additional pathways to reduce carbon footprints. By releasing their code and detailed experimental pipeline, the authors promote reproducibility and encourage the community to adopt environmentally conscious evaluation practices.

In summary, this study provides the first comprehensive quantification of carbon emissions for GNN‑based recommender systems, highlights embedding size as a decisive factor in environmental impact, and offers actionable guidance for researchers and practitioners seeking to develop high‑performing yet sustainable recommendation technologies.