Abstract
Multimodal Recommender Systems (MRSs) aim to provide better suggestions by combining user behavior with item content. However, they are often hindered by the "tug-of-war" between collaborative and modal signals and the inability to capture broad user interests from sparse data. We present LGMRec, which decouples local graph learning and introduces a global hypergraph module to model implicit attributes. LGMRec achieves state-of-the-art performance on Amazon benchmark datasets, particularly for users with few interactions.
Introduction
Multimodal Recommender Systems (MRSs) face two critical bottlenecks that limit their effectiveness:
- The Coupling Problem: Most models share a single "User ID" embedding across both collaborative filtering and multimodal modules. Empirical analysis reveals a "gradient conflict": in over 50% of training cases, the collaborative signal and the visual/textual signals provide contradictory guidance during optimization. This "tug-of-war" prevents the model from converging to an optimal state.
- The Locality Problem: Existing graph-based models focus on "local" neighbors—items a user has directly interacted with. In sparse data scenarios, the model misses "global" interests, such as a general preference for a specific style across different categories.
Methodology
LGMRec introduces a dual-pronged architecture to solve these issues:
Local Graph Embedding (LGE) Module
To solve the coupling problem, LGE independently learns collaborative-related and modality-related embeddings. Instead of forcing one ID embedding to perform multiple roles, it uses separate message-passing tracks. This ensures that the collaborative signal is not muddied by modal signals at the local level.
Global Hypergraph Embedding (GHE) Module
To capture broader interests, LGMRec utilizes hypergraphs. Hyperedges represent "implicit attributes" (like style or color). Items with similar visual or textual features are grouped into the same hyperedge, allowing the model to learn global dependencies that are not visible in the sparse local interaction graph.
Fusion and Contrastive Learning
The model uses a cross-modal hypergraph contrastive loss (InfoNCE) to align the global representations learned from different modalities. Finally, these decoupled local and global embeddings are fused using an adjustable factor ($\alpha$) to produce the final recommendation score.
Experimental Results
The authors evaluated LGMRec on three Amazon benchmark datasets: Baby, Sports, and Clothing.
- Performance Gains: LGMRec consistently outperformed state-of-the-art baselines like LightGCN and MMGCN. On the Baby dataset, it showed a 12.98% improvement in Recall@10.
- Sparsity Champion: In stress tests for data sparsity, LGMRec's performance gap over other models increased as data became sparser. For the most sparse user group in the Baby dataset, it realized a 19.95% improvement over the next best model.
- Ablation Success: Removing either the Global (GHE) or Local (LGE) modules led to performance drops, confirming that both "zoomed-in" and "zoomed-out" views are necessary.
Conclusion
LGMRec addresses fundamental flaws in multimodal recommendation by decoupling conflicting signals and modeling global dependencies. While hypergraph learning introduces some computational overhead, the significant gains in accuracy and robustness to data sparsity make it a compelling approach for modern recommendation engines.
Summary updated