Uncertain Multi-Objective Recommendation via Orthogonal Meta-Learning Enhanced Bayesian Optimization

Hongxu Wang, Zhu Sun, Yingpeng Du, Lu Zhang, Tiantian He, Yew-Soon OngarXiv 2025 · DOI · arXiv

Introduction

Modern recommender systems (RS) are often criticized for being "accuracy-obsessed," leading to filter bubbles and neglecting ethical dimensions like diversity and fairness. While multi-objective optimization exists, it typically relies on fixed trade-off weights that fail to account for individual user preferences or the non-linear relationship between weights and outcomes. This paper introduces BOOML (Bayesian Optimization boosted via Orthogonal Meta-Learning), a framework designed to achieve "Level 4 Autonomy" in RS by automatically discovering personalized objective balances.

The Challenge: Moving Beyond Fixed Objectives

Traditional RS operate at lower levels of autonomy, relying on human-defined rules or simple profile-based accuracy. The authors argue for a shift toward Level 4 Autonomy, where systems independently identify and optimize uncertain, multi-objective environments.

Levels of autonomy in recommender systems, progressing from rule-based to uncertain multi-objective optimization.

In these complex environments, the relationship between hyperparameter weights (e.g., for diversity or fairness) and final performance is a "black box." Simply increasing a weight does not guarantee a linear improvement in that metric, necessitating a more sophisticated optimization strategy.

Methodology: BOOML

BOOML addresses these challenges through three core components:

  1. Bayesian Optimization (BO): BO is employed to search the weight space efficiently. By modeling the performance as a Gaussian process, the system can identify optimal trade-offs without exhaustive grid searches.
  2. Group-Level Personalization: Recognizing that individual-level BO is computationally prohibitive, the framework clusters users into groups based on historical behavior, finding optimal weights for each cluster.
  3. Orthogonal Meta-Learning: To make the iterative BO process feasible, the authors use meta-learning to accelerate model convergence. Furthermore, they implement Orthogonal Gradient Descent (PCGrad) to ensure that conflicting objectives—like accuracy and diversity—do not cancel each other out during training.
Training dynamics of BOOML, showing total loss convergence and the adaptive balancing of objective weights over epochs.

Experimental Evaluation

The framework was tested on three Amazon datasets (Games, Electronics, and Movies). Key findings include:

  • Efficiency: Meta-learning reduced training time from 40-80 epochs to just 5, making the BO approach practical.
  • Performance: BOOML outperformed state-of-the-art baselines (SMORL, MMoE) on aggregate metrics like Rescaled Sum and Harmonic Mean.
  • Accuracy Retention: Unlike many models that sacrifice accuracy for diversity, BOOML often improved NDCG while simultaneously enhancing ethical metrics.

Conclusion

BOOML represents a significant step toward autonomous, ethical recommender systems. By treating objective balancing as a black-box optimization problem and leveraging meta-learning for efficiency, it provides a scalable way to deliver personalized, diverse, and fair recommendations. Future work will explore temporal dynamics and broader ethical dimensions like privacy and transparency.

Summary updated

Abstract

Recommender systems (RSs) play a crucial role in shaping our digital interactions, influencing how we access and engage with information across various domains. Traditional research has predominantly centered on maximizing recommendation accuracy, often leading to unintended side effects such as echo chambers and constrained user experiences. Drawing inspiration from autonomous driving, we introduce a novel framework that categorizes RS autonomy into five distinct levels, ranging from basic rule-based accuracy-driven systems to behavior-aware, uncertain multi-objective RSs - where users may have varying needs, such as accuracy, diversity, and fairness. In response, we propose an approach that dynamically identifies and optimizes multiple objectives based on individual user preferences, fostering more ethical and intelligent user-centric recommendations. To navigate the uncertainty inherent in multi-objective RSs, we develop a Bayesian optimization (BO) framework that captures personalized trade-offs between different objectives while accounting for their uncertain interdependencies. Furthermore, we introduce an orthogonal meta-learning paradigm to enhance BO efficiency and effectiveness by leveraging shared knowledge across similar tasks and mitigating conflicts among objectives through the discovery of orthogonal information. Finally, extensive empirical evaluations demonstrate the effectiveness of our method in optimizing uncertain multi-objectives for individual users, paving the way for more adaptive and user-focused RSs.
BibTeX
@misc{wang2025uncertain,
  author = {Hongxu Wang and Zhu Sun and Yingpeng Du and Lu Zhang and Tiantian He and Yew-Soon Ong},
  title = {{Uncertain Multi-Objective Recommendation via Orthogonal Meta-Learning Enhanced Bayesian Optimization}},
  year = {2025},
  eprint = {2502.13180},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.13180},
}
Figure