RoFiGR: Coarse Routing, Fine Refinement via Level-Adaptive Distillation in Multimodal Generative Recommendation
What it is
RoFiGR is a two-tree generative recommender that improves multimodal recommendation by aligning semantic identifier (SID) codes with collaborative item embeddings. It uses a sequential teacher to supervise legal branches during generation, employing fixed coarse-level weights and a bounded adaptive residual to balance routing and fine-grained prediction.
The problem it solves
Generative recommenders predict semantic identifiers sequentially, so early routing errors can exclude relevant items before fine-grained prediction begins; existing methods often fail to adequately incorporate collaborative filtering signals into these coarse decisions.
What you could use it for
- Multimodal recommendation systems where early routing errors in semantic identifiers can exclude relevant items
- Improving ranking metrics (NDCG, Recall) in generative recommenders by incorporating collaborative filtering signals into SID construction
- Distillation-based training for hierarchical generative models
Why it matters
If successful, RoFiGR could significantly improve the accuracy and robustness of generative recommenders by ensuring that coarse routing decisions reflect both item semantics and user collaborative preferences, reducing the risk of early exclusion of relevant candidates.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.