Towards Sustainability-aware Recommender Systems: Analyzing the Trade-of Between Algorithms Performance and Carbon Footprint Giuseppe Spillo giuseppe.spillo@uniba.it University of Bari Aldo Moro Bari, Italy Allegra De Filippo allegra.deflippo@unibo.it University of Bologna Bologna, Italy Cataldo Musto cataldo.musto@uniba.it University of Bari Aldo Moro Bari, Italy Michela Milano michela.milano@unibo.it University of Bologna Bologna, Italy Giovanni Semeraro giovanni.semeraro@uniba.it University of Bari Aldo Moro Bari, Italy ABSTRACT In this paper, we present a comparative analysis of the trade-of between the performance of state-of-the-art recommendation algo- rithms and their environmental impact. In particular, we compared 18 popular recommendation algorithms in terms of both perfor- mance metrics (i.e., accuracy and diversity of the recommendations) as well as in terms of energy consumption and carbon footprint on three diferent datasets. In order to obtain a fair comparison, all the algorithms were run based on the implementations available in a popular recommendation library, i.e., RecBole, and used the same experimental settings. The outcomes of the experiments showed that the choice of the optimal recommendation algorithm requires a thorough analysis, since more sophisticated algorithms often led to tiny improvements at the cost of an exponential increase of carbon emissions. Through this paper, we aim to shed light on the prob- lem of carbon footprint and energy consumption of recommender systems, and we make the frst step towards the development of sustainability-aware recommendation algorithms. CCS CONCEPTS · Information systems → Recommender systems. KEYWORDS recommender systems, evaluation, sustainability, non-accuracy metrics, carbon footprint ACM Reference Format: Giuseppe Spillo, Allegra De Filippo, Cataldo Musto, Michela Milano, and Gio- vanni Semeraro. 2023. Towards Sustainability-aware Recommender Systems: Analyzing the Trade-of Between Algorithms Performance and Carbon Foot- print. In Seventeenth ACM Conference on Recommender Systems (RecSys ’23), Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proft or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specifc permission and/or a fee. Request permissions from permissions@acm.org. RecSys ’23, September 18ś22, 2023, Singapore, Singapore © 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 979-8-4007-0241-9/23/09. . . $15.00 https://doi.org/10.1145/3604915.3608840 September 18ś22, 2023, Singapore, Singapore. ACM, New York, NY, USA, 7 pages. https://doi.org/10.1145/3604915.3608840 1 BACKGROUND AND MOTIVATIONS Artifcial Intelligence (AI) algorithms have made signifcant strides in recent years, achieving state-of-the-art performance in computer vision, natural language processing (NLP), robotics, and game play- ing. However, all these advances came at the cost of a huge increase in terms of energy consumption and carbon emissions. This makes AI being considered as a power-hungry technology [34], since the computational demands of these algorithms typically require huge computing power and long training time, which in turn imply a large energy consumption. Indeed, back in 2018, some studies have estimated that the energy usage of data centers already represented close to 1% of the global energy usage [15]. Despite the growing awareness of researchers and practitioners in topics related to en- ergy efciency of AI algorithms [4, 5], the environmental costs asso- ciated with large-scale computation, and the trade-of between the performance of AI algorithms and their sustainability in terms of carbon footprint are not being fully captured by current studies. The problem is particularly relevant, since the concentration of greenhouse gases (GHGs) in the atmosphere has a dramatic infuence on climate change [13]. As an example, Strubell et al. [26] found that training recent translation engines once can emit up 280 tonnes of  2 , which is equivalent to the energy use of 35 homes in one year 1 . For this reason, it is important to estimate and curtail both the energy used and the emissions produced by training and deploying AI models [25], in order to address the sustainability of developing and using AI systems [29]. Generally speaking, this research is framed in the general area of sustainability of AI, which deals with measuring the sustainability of developing and using AI models. However, it is important to point out that the concept of sustainable AI is very broad, since it is based on three fundamental pillars [23]: social, economic and envi- ronmental. In this paper, we just focus on the latter pillar, since we analyze the environmental impact of AI (specifcally, recommenda- tion) algorithms. This includes the selection of the right amount of data that are needed to train a model as well as the identifcation of 1 Estimation based on: https://www.epa.gov/energy/greenhouse-gas-equivalencies- calculator 856