AMORe: An Agnostic Multi-Objective Framework for Recommendation
Recommender systems increasingly need to balance relevance with beyond-accuracy objectives, including fairness and long-tail exposure. Existing multi-objective approaches often rely on objective-specific loss designs, which limits reuse when target metrics or their numerical scales change. We propose AMORe, a scale-aware framework for optimizing multiple decomposable top- k recommendation objectives within gradient-based recommender training. AMORe combines differentiable ranking approximators with a lightweight top- k smoothing component and a unified loss that minimizes normalized deviations from user-specified utopia points. Experiments on four public datasets and three recommendation backbones show that AMORe exposes controllable trade-offs between relevance and beyond-accuracy objectives, with improvements in beyond-accuracy metrics over corresponding vanilla backbones in most evaluated settings. Comparisons with representative fairness-aware baselines, together with ablation and efficiency analyses, clarify when AMORe improves trade-off stability and quantify the computational overhead introduced by multi-objective training. We release code and datasets at: https://github.com/vincpapa/Agnostic-Multi-Objective-Recommendation .
Authors
- Tommaso Di Noia (ORCID: https://orcid.org/0000-0002-0939-5462)
- Raffaele Perego (ORCID: https://orcid.org/0000-0001-7189-4724)
- Vito Walter Anelli (ORCID: https://orcid.org/0000-0002-5567-4307)
- Franco Maria Nardini (ORCID: https://orcid.org/0000-0003-3183-334X)
- Vincenzo Paparella (ORCID: https://orcid.org/0000-0002-5602-1682)
Institutions
- Istituto di Scienza e Tecnologie dell'Informazione "Alessandro Faedo" (IT)
- Polytechnic University of Bari (IT)
Publication Details
- Journal
- ACM Transactions on Information Systems
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1145/3857349
- Primary Topic
- Recommender Systems and Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00