A Narrative Review of Multi-Criteria Recommender Systems: from Perspectives of Multi-Criteria Decision Making

Traditional recommender systems primarily rely on overall ratings, which may fail to capture the fine-grained nuances of user preferences. Multi-Criteria Recommender Systems (MCRS) address this limitation by incorporating user feedback across multiple aspects of the items to provide more accurate and personalized suggestions. While the integration of advanced deep learning techniques has significantly improved MCRS performance, the research landscape remains fragmented, lacking a unified theoretical framework that connects modern neural network architectures with classical Multi-Criteria Decision-Making (MCDM) theories. To address this gap, this paper presents a conceptually driven narrative review that introduces a novel architectural taxonomy for MCRS. We systematically classify existing methodologies into two paradigms: Multi-Stage architectures and Single-Stage architectures. Multi-Stage MCRS explicitly integrates MCDM principles, such as multi-attribute utility functions or Pareto dominance, by decomposing the recommendation pipeline into multi-criteria rating prediction and preference aggregation. Conversely, Single-Stage MCRS employs holistic, end-to-end modeling paradigms, such as neighborhood-based methods, deep tensor factorization, or multiview graph neural networks, in order to capture complex and non-linear dimensional interactions simultaneously. Moreover, this review examines the transition from relying on explicit criteria elicitation to mining implicit preferences from user-generated reviews, a shift that significantly broadens the real-world applicability of MCRS. We also discuss ongoing challenges in the field, such as data sparsity, integration with contextual situations, and model explainability. By bridging classical decision science with modern deep learning, this review provides a clear roadmap and theoretical insights for future MCRS research.