Introduction
Traditional recommender systems typically rely on a single overall rating—the classic "five-star" score. However, human preferences are inherently multi-faceted; a user might appreciate a movie's visual effects while finding its narrative lacking. Multi-Criteria Recommender Systems (MCRS) aim to capture these nuances to provide more accurate suggestions.
This paper introduces D-MGAC, a novel framework that leverages multiview graph structures, dual-attention mechanisms, and contrastive learning to model complex multi-criteria feedback. By moving beyond single-dimension ratings, D-MGAC captures both specific niche preferences and broad global trends, significantly outperforming traditional recommendation baselines.
The Challenge of Multi-Criteria Recommendation
While MCRS offer a more detailed view of user preferences, they face two significant hurdles: data sparsity and complex relationship modeling. Most users only rate a small fraction of available items, leading to sparse interaction matrices. Furthermore, existing models often struggle to balance "local" relationships (behavior within a specific criterion like 'price') with "global" relationships (how different criteria interact to form an overall preference).
As illustrated above, a multi-criteria system introduces a layer of complexity by mapping user-item interactions across multiple distinct criteria, requiring more sophisticated modeling techniques than standard bipartite graphs.
The D-MGAC Framework
The proposed D-MGAC (Dual Multiview and Graph Attention and Contrastive Learning) framework addresses these challenges through a multi-layered architecture:
- Multi-Edge Bipartite Graph: The system represents user-item interactions using multiple edges, where each edge corresponds to a specific criterion rating.
- Multiview Dual Graph Attention (MDGAT): This component employs Local Attention to identify important neighbors within a single criterion and Global Attention to weigh node importance across the entire multi-criteria landscape. This dual approach captures long-range dependencies often missed by single-view models.
- Dual Contrastive Learning: To mitigate data sparsity, the model uses self-supervised learning. It identifies "anchor points" based on neighborhood similarity and trains the model to align similar user/item representations while separating dissimilar ones, both locally and globally.
- Embedding Fusion: Criterion-specific embeddings are concatenated into a unified vector, which is then used to predict ratings via a Support Vector Machine (SVM).
Experimental Evaluation
The model was evaluated against 12 baseline algorithms, including Matrix Factorization (BMF) and Deep Multi-Criteria Collaborative Filtering (DMCF), across three real-world datasets:
- Yahoo!Movies: Includes ratings for story, acting, direction, and visuals. D-MGAC achieved a Mean Absolute Error (MAE) of 0.6105, outperforming the best baseline (CFM-item at 0.6127).
- BeerAdvocate: Includes ratings for aroma, appearance, palate, and taste. D-MGAC achieved an MAE of 0.4156, significantly better than the closest baseline (BMF at 0.4394).
- TripAdvisor: A high-sparsity dataset (99.9%) used to test robustness across seven criteria (e.g., service, value, sleep quality).
Key Insights and Results
- The Global Advantage: Ablation studies demonstrated that global attention is essential for capturing the "big picture" across all criteria. Removing this component (D-MGAC*) led to a visible drop in performance.
- Sparsity Resilience: Contrastive learning acted as a powerful denoiser, allowing the model to generate meaningful recommendations even when the interaction matrix was mostly empty. The use of similarity-based "anchor points" rather than random sampling provided more stable training.
- Complexity Trade-offs: While adding more criteria generally improves accuracy, it also increases computational overhead. The dual-attention mechanism, while effective, is computationally intensive.
Conclusion and Future Work
D-MGAC represents a significant step forward in multi-criteria recommendation by effectively combining multiview graph attention with contrastive learning. Future research will focus on addressing computational scalability for massive datasets and incorporating "temporal drift" to account for how user preferences evolve over time. Additionally, exploring Dynamic Hypergraph Learning could help capture even more complex, non-pairwise relationships between users and items.
Summary updated