Sustainable and Health Aware Recommendation System (SHARS)
Singh, Udham (2026)
2026
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-202604176766
https://urn.fi/URN:NBN:fi:amk-202604176766
Tiivistelmä
Traditional food recommendation systems (FRS) primarily prioritize predictive accuracy, often overlooking dietary health and environmental sustainability. This thesis presents SHARS (Sustainable and Health-Aware Recommendation System), a novel three-phase framework developed to mitigate the global ‘Food-Planet-Health’ trilemma.
The research objectives were addressed through a three-phase architecture comprising collaborative pre-training, hybrid fine-tuning, and multi-objective ranking. Collaborative embeddings were learned via Neural Matrix Factorization and refined using a Hybrid Gating Network integrated with 813 content-based dimensions. A multi-objective optimization (MOO) ranking was then executed to balance user preference, health, water footprint, and carbon footprint.
Empirical results demonstrated that SHARS achieved a Huber loss of 0.3681 and an RMSE of 0.6067, significantly outperforming the collaborative filtering baseline. Substantial ecological improvements were recorded, including a 68.6% reduction in carbon emissions and a 59.3% reduction in water usage. Furthermore, health scores were increased by 57.8% for warm-start users, while environmental efficacy was maintained for cold-start users with carbon and water reductions of 59.6% and 58.8%, respectively. User preference was stabilized at a score of 6.00, validating that recommendation utility was not compromised by sustainability constraints.
Integrating sustainability and health 'nudges' does not compromise recommendation utility. Consequently, this framework serves as a scalable technical blueprint for the development of multi-dimensional recommender systems that align individual preferences with global ecological imperatives.
The research objectives were addressed through a three-phase architecture comprising collaborative pre-training, hybrid fine-tuning, and multi-objective ranking. Collaborative embeddings were learned via Neural Matrix Factorization and refined using a Hybrid Gating Network integrated with 813 content-based dimensions. A multi-objective optimization (MOO) ranking was then executed to balance user preference, health, water footprint, and carbon footprint.
Empirical results demonstrated that SHARS achieved a Huber loss of 0.3681 and an RMSE of 0.6067, significantly outperforming the collaborative filtering baseline. Substantial ecological improvements were recorded, including a 68.6% reduction in carbon emissions and a 59.3% reduction in water usage. Furthermore, health scores were increased by 57.8% for warm-start users, while environmental efficacy was maintained for cold-start users with carbon and water reductions of 59.6% and 58.8%, respectively. User preference was stabilized at a score of 6.00, validating that recommendation utility was not compromised by sustainability constraints.
Integrating sustainability and health 'nudges' does not compromise recommendation utility. Consequently, this framework serves as a scalable technical blueprint for the development of multi-dimensional recommender systems that align individual preferences with global ecological imperatives.