GRACE fine-tunes recommenders for eco-friendly choices without retraining

A team from multiple universities (Yibowen Zhao, Yinan Zhang, Ning Liu, Lizhen Cui, Chunyan Miao) has introduced GRACE (Green Recommendation via Adaptive Conflict-rEsolution), a fine-tuning framework designed to make recommender systems more environmentally sustainable without sacrificing personalization accuracy. Traditional approaches to "green" recommendation either train new models from scratch—consuming significant computational energy—or add a reranking step at inference that increases latency. GRACE sidesteps both costs by fine-tuning existing pretrained models with item-level sustainability signals such as eco-scores or health indices. A key innovation is its use of a differentiable approximation for discrete, non-differentiable green values, which enables direct optimization of the sustainability objective rather than relying on pairwise comparisons.

To maintain personalization quality, GRACE employs a gradient projection mechanism that mediates conflicts between the green objective and accuracy objective during fine-tuning. This ensures that steering users toward eco-friendlier choices does not drastically degrade the relevance of recommendations. Experiments on real-world datasets show that GRACE improves sustainability-oriented outcomes while generally preserving recommendation accuracy, thanks to a controllable preference-anchored update mechanism. The approach is particularly promising for platforms that want to reduce carbon footprints without building new infrastructure or increasing inference costs. As sustainability becomes a priority in AI, GRACE offers a practical path to greener recommendation at minimal overhead.