The Environmental Cost of Recommender Systems

Abstract

As global warming intensifies, the environmental impact of research is increasingly scrutinized, yet the carbon footprint of recommender systems remains largely unexplored. This talk delves into our recent analysis presented at ACM RecSys 2024, where we measured the carbon emissions of 79 papers from RecSys 2013 and 2023. By reproducing typical experimental pipelines, we found that deep learning-based recommender systems emit, on average, 42 times more CO2 equivalents than traditional algorithms, contributing significantly to the field’s overall environmental toll. We advocate for greater awareness and adoption of sustainable practices in recommender systems research to mitigate this impact.