Sustainability is a complex topic, and it needs consideration in all aspects of recommendations. For example, the fashion industry's significant environmental and social impact necessitates new approaches to promote sustainable consumption behaviors. This work focuses specifically on two critical and interconnected aspects of sustainable recommendation: promoting environmentally conscious user behavior and reducing the computational environmental footprint of recommendation algorithms. First, it investigates the application of digital nudging techniques in fashion recommender systems to encourage environmentally conscious purchasing decisions. It describes a comprehensive user study with 251 participants, testing three interface designs that incorporate textual nudges and visual sustainability labels. Results demonstrate remarkable success in promoting second-hand garment selection, with 50-67% of participants choosing used items compared to a 7% market rate. However, the computational requirements of multimodal recommendation systems introduce their own environmental considerations. To address this paradox, this work describes a Transformer-based architecture leveraging attention bottlenecks for more eficient multimodal fusion. Early experiments suggest a significant reduction in representation dimensionality while maintaining recommendation performance. The main goal of this work is to provide a complete picture of what sustainability in recommendation entails, from consumer decisions to resource usage by recommendation algorithms.
CEUR
ceur-ws.org while ofering enhanced accuracy through rich data representations, demand substantial computational resources for both training and inference. Section 4 describes our approach to reduce this impact.
Starting from these approaches, the goal is their integration as a step toward holistic, sustainable recommendation systems that consider both the outcomes they promote and the resources they consume in achieving those outcomes.
The concept of nudging, originally developed in behavioral economics, involves modifying subtle
aspects within choice environments to encourage specific decisions while preserving individual freedom
of choice. Weinmann et al. defined digital nudging as “the use of user-interface design elements to guide
people’s behavior in digital choice environments” [
As a test bed for nudging, we focused on fashion, as it presents a particularly interesting case study
for digital nudging applications due to its significant environmental and social footprint. Promoting
sustainable fashion consumption is particularly challenging, as several complex factors influence buying
decisions. Fashion serves as a symbol of personal identity and social status [
To investigate the efectiveness of digital nudging in promoting sustainable fashion consumption, we
conducted a comprehensive user study involving 251 participants [
LABELSH recommended only second-hand garments but combined textual nudges with visual labels summarizing sustainability and ethical standards, including detailed information about CO₂ emissions, water consumption, workers’ well-being, and respect for animals. (iii) LABELSHNEW presented both new and second-hand alternatives with the same visual labels as LABELSH.
The results of the experiment showed remarkable success in promoting second-hand garment
selection. Across the diferent user interfaces, 50-67% of participants chose second-hand products as
their final selection, a dramatic improvement over the 7% estimated from previous market studies [
After interacting with each user interface, we asked users to evaluate the decision-making process, the interface adequacy and their satisfaction. The LABELSH interface received the highest ratings for decision-making support, with users finding the visual labels particularly useful for product comparison. Conversely, the LABELSHNEW interface was perceived as the most intuitive and informative, benefiting from its comprehensive presentation of product types. It also received the best overall satisfaction score.
Additionally, we asked users to pinpoint the factors that influenced their choices. While price was the most influential (ranging from 56 to 59%, depending on the interface), as well as style (39-49%), color (60-64%), and materials (43-53%), users also frequently considered environmental sustainability (33-38%) and ethical standards (38-40%). This is in contrast with the data from final selections above, which show that environmental sustainability is more influential in practice than ethical standards.
The preliminary results we obtained suggest increasing the adoption of nudges in clothes recommender systems to enhance user awareness about items, their sustainability, and their social impact.
The environmental impact of digital systems has gained increasing attention as data centers and machine learning operations consume growing amounts of energy.
While digital nudging can promote sustainable consumption behaviors, the computational infrastructure supporting recommender systems introduces its own environmental considerations. Modern systems, particularly those utilizing multimodal item representations (incorporating text, images, audio, and other data types), require substantial computational resources for training and operation.
Multimodal recommendation algorithms often rely on very high-dimensional data. For example, in the context of an e-commerce website for fashion, the following information could be fed as input to the algorithm: (i) Visual features: High-resolution image embeddings capturing style, color, and aesthetic properties; (ii) Textual descriptions: Natural language processing of product descriptions, materials, and brand information; (iii) User interaction data: Historical preferences and behavioral patterns.
While these rich representations enable more accurate and nuanced recommendations, they also
create computational challenges. The high dimensionality of multimodal features introduces significant
overhead in terms of memory usage, processing time, and energy consumption during both training
and inference phases. At the same time, however, diferent modalities could share some information,
resulting in redundancy. As a high-level example, a garment’s color might be present both in the
visual embeddings from the product image, and in the textual description. Information fusion involves
combining multiple information sources into one [
Traditional approaches to multimodal fusion have been applied to recommendation in the literature
[
We are currently experimenting with a novel Transformer-based architecture leveraging attention bottlenecks [18] to develop a plug-and-play approach for more eficient multimodal recommendation, which is agnostic to the specific algorithm used. Early experiments suggest that this approach can significantly reduce representation dimensionality while maintaining or even improving recommendation performance across multiple algorithms and datasets.
The above findings on digital nudging efectiveness and computational eficiency challenges point toward an important future direction, i.e., the development of comprehensive, sustainable recommendation systems that integrate user-centered sustainability promotion and resource-eficient algorithmic design. Rather than treating these as separate concerns, future research should focus on creating unified frameworks that optimize simultaneously for sustainable user behavior and minimal environmental impact. For example, we plan to further develop our fusion mechanism, making it possible to prioritize salient features such as sustainability and ethics, enabling seamless integration with the user interface for the nudging we developed.
Furthermore, using the data collected in [
The ultimate vision encompasses recommender systems that serve as comprehensive sustainability platforms, integrating eficient algorithmic design with efective behavioral interventions to create measurable positive environmental impact while maintaining user satisfaction and commercial viability.
I want to thank my supervisors, Liliana Ardissono and Noemi Mauro, for their invaluable guidance and support, which has been essential in developing this work.
During the preparation of this work, the author used Claude2 to: Paraphrase and reword, and Grammar and spelling check. After using these tools, the author reviewed and edited the content as needed. The author takes full responsibility for the publication’s content. Context-Aware Recommendations, in: Fourteenth ACM Conference on Recommender Systems, ACM, Virtual Event Brazil, 2020, pp. 338–347. URL: https://dl.acm.org/doi/10.1145/3383313.3412268. doi:10.1145/3383313.3412268. [17] J. Liu, D. Capurro, A. Nguyen, K. Verspoor, Attention-based multimodal fusion with contrast for robust clinical prediction in the face of missing modalities, Journal of Biomedical Informatics 145 (2023) 104466. URL: https://linkinghub.elsevier.com/retrieve/pii/S1532046423001879. doi:10.1016/ j.jbi.2023.104466. [18] A. Nagrani, S. Yang, A. Arnab, A. Jansen, C. Schmid, C. Sun, Attention Bottlenecks for Multimodal Fusion, in: In Proceedings of the 35th Conference on Neural Information Processing Systems (NeurIPS 2021)., 2021. URL: https://proceedings.neurips.cc/paper/2021/file/ 76ba9f564ebbc35b1014ac498fafadd0-Paper.pdf.