.png){kind=icon}
.png){kind=icon}
Return
An adaptive scalarization framework for multi-objective recommender systems
.png){kind=icon}
PRE
AI
DOI：10.1007/s10115-026-02828-9.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recommender systems increasingly face the challenge of balancing competing objectives such as accuracy, fairness, and novelty. Traditional approaches often optimize a single objective, leading to biased outcomes, limited content exposure, and reduced user satisfaction. This challenge becomes more complex when system priorities evolve over time or vary across user groups. To address this, we propose FairRec-ARPM, an adaptive scalarization-based framework that jointly optimizes multiple objectives through a responsive and interpretable strategy. The method introduces a reference point update mechanism to adjust optimization based on performance gaps, along with a priority-aware scalarization function that emphasizes underperforming objectives while maintaining overall balance, without relying on fixed preferences or manual tuning. The framework integrates collaborative filtering with demographic fairness profiling in a unified optimization loop, enabling ethical and personalized recommendations. Experiments on four benchmark datasets (MovieLens 1M, RetailRocket, Amazon, and Yelp) show that FairRec-ARPM consistently outperforms strong baselines, achieving improvements of up to 6.6% in Recall@10, 13.1% in Demographic Parity, 3.8% in Novelty, and 13.7% in Diversity, while reducing runtime by up to 3 $$\times $$ compared to Pareto-based methods. These results demonstrate the effectiveness of FairRec-ARPM in managing multi-objective trade-offs, making it suitable for real-world recommender systems where fairness and personalization must coexist efficiently and transparently.
Keywords:
Multi-objective recommender systems
Adaptive scalarization
Fairness-aware recommendation
Reference point optimization
Personalized recommendation
Trade-off recommendation
Journal
IF:
3.1
Papers:
517
Citations:
5.2K