A Holistic View of Sustainability in Multimodal Recommender Systems

Sustainability is a complex topic that must be considered in all aspects of recommendations, including what is recommended and which technologies are used for item suggestion. For example, the fashion industry’s significant environmental and social impact demands new approaches to promote sustainable consumption behaviors. However, the environmental impact of the applications supporting information exploration cannot be neglected, as their training and execution impact the environment. Previously, we investigated the application of digital nudging techniques to encourage environmentally conscious fashion consumption, focusing on user satisfaction and conversion rates. In this paper, we focus on the energy and computational requirements of multimodal recommendation systems, which underlie the personalized suggestion of items in current product catalogs. We present the preliminary results of an experiment comparing the resource, energy, and time consumption of a set of state-of-the-art multimodal recommendation algorithms, considering their training and testing phases. For this experiment, we used two Amazon datasets (one about clothing, the other about sports products). The experimental results reveal an important trade-off between the recommendation performance and the environmental sustainability of the algorithms. Among those we compared, CVAE and CDL appeared to be preferable on a single dataset. However, neither of them outperformed the other algorithms in all the evaluation dimensions we considered. These findings suggest the need for work to develop recommendation algorithms that combine resource efficiency with recommendation capability. Work is also needed to better integrate recommendation algorithms with user interfaces, highlighting the environmental impact of products, to steer sustainable product consumption.