Rethinking ML Model Selection Using Sustainable HCI

Jury

• Aurelie BUGEAU, Professor, Université de Bordeaux, Rapporteure & Examinatrice
• Ines DI LORETO, Associate Professor, Université de technologie de Troyes, Rapporteure & Examinatrice
• Aurélien TABARD, Maitre de conférences, Université Claude Bernard Lyon, Examinateur
• Anne-Laure LIGOZAT, Professor, Université Paris-Saclay et École Nationale Supérieure d’Informatique pour l’Industrie et l’Entreprise (ENSIIE), Examinatrice

Abstract

Training and deploying large models demand substantial computational resources, contribute to carbon emissions, and deepen inequalities in access to data and infrastructure. Yet, sustainability is still rarely treated as a primary selection requirement, often being ignored in front of performance and efficiency goals. A crucial but largely ignored decision-making point in this process is model selection, where developers choose between competing models, often prioritizing accuracy or familiarity over efficiency and sustainability. Human–Computer Interaction (HCI) focuses on interactions between people, technology, and context, and offers valuable perspectives for addressing these challenges. By emphasizing human values and reflection into the model selection process, HCI methods can help practitioners consider the environmental consequences of their technical choices.

This thesis explores how sustainability can be included into ML workflows by combining perspectives from HCI, ML, and sustainability research. It includes theoretical, empirical, and technical contributions to advance understanding and practice toward more sustainable and human-centered machine learning. I examine the current state of sustainability awareness and considerations within HCI and ML through interviews, questionnaires, and a comprehensive literature review. Building on these insights, I introduce a new framework for sustainability research that connects both fields and outlines future research directions across their intersection. Through interviews with ML practitioners, I identified how initial problem definition and model selection is critical for reducing the sustainability impact of AI applications. My results show that sustainability awareness remains limited and often secondary to performance goals when selecting a model. Although developers are aware of the energy and infrastructural costs associated with large ML models, they lack tools or frameworks to evaluate or reduce these effects. 

These findings resulted in the design and evaluation of Seleco, an interactive interface that guides developers in defining their ML tasks and selecting suitable models while considering their energy consumption and environmental impact. Our comparative structured observation study with professional ML developers shows that providing transparency about model energy use and environmental footprint encourages them to reflect on their choices and to consider sustainability as part of their everyday decision-making. Beyond the tool itself, my findings demonstrate how applying HCI methods to ML processes can promote sustainable practices and encourage practitioners to better navigate the trade-offs between technical performance and ethical or environmental responsibility. In conclusion this thesis demonstrates how integrating human-centered design, reflective practice, and sustainability considerations can foster more responsible and transparent approaches to ML development.