Investigating Carbon Footprint of Recommender Systems Beyond Training Time

Abstract

This study extends environmental evaluation of recommender systems beyond training by incorporating inference-time emissions and examining how training configurations affect total carbon footprint. Results show that models with higher training emissions can sometimes have lower environmental costs during prolonged inference, and that reducing unnecessary validation-metric computation can meaningfully lower overall emissions.

Publication
_Proceedings of the Nineteenth ACM Conference on Recommender Systems (RecSys 2025), pp. 1206-1211, https://doi.org/10.1145/3705328.3759324_