In-processing and post-processing strategies for balancing accuracy and sustainability in product recommendations

Abstract

Many e-commerce websites use product recommendation systems. With the growing awareness of the environmental impact of our consumption, these recommender systems, well-known for encouraging purchases and consumption, are being challenged. In recent literature, it has been suggested that recommender systems should balance the exploitation of existing preferences with the exploration of sustainable items, i.e., to make sustainable alternatives more accessible to consumers and promote sustainable consumption habits. In this paper, we therefore analyse how to increase the presence of sustainable products in recommendation lists, without overly decreasing their accuracy. More precisely, we test three in-processing and four post-processing strategies using an offline experimental design. The post-processing strategies 1 (relevance scores’ adjustment) and 3.1. (incremental list formation with calibration) manage to offer interesting accuracy-sustainability trade-offs on our datasets. For instance, by applying post-processing strategy 3.1 to a content-based recommendation algorithm, a gain of up to 20% can be achieved for the sustainability metric without any loss of accuracy. Greater sustainability improvements can be achieved if a loss of accuracy is tolerated. For practitioners, i.e., e-commerce platforms, this means that they could continue to offer relevant recommendations while promoting a more sustainable consumption.