Learning Multi-aspect Item Palette: A Semantic Tokenization Framework for Generative Recommendation

Traditional recommendation models rely on item IDs, limiting their use of item content and generalization to long-tailed or cold-start items. Semantic tokenization, which converts item semantics into discrete tokens, is a promising approach for generative recommendation models. However, existing RQ-VAE-based methods face challenges with embedding extraction, hierarchical quantization, and training stability. To overcome these limitations, we propose LAMIA, a novel multi-aspect semantic tokenization approach that learns an “item palette”—a set of independent, semantically parallel embeddings capturing diverse item aspects. LAMIA further improves semantic encoders through domain-specific tuning with text-based reconstruction tasks, yielding more representative embeddings. Extensive experiments show that LAMIA significantly outperforms existing methods in recommendation accuracy across multiple tasks and datasets. For reproducibility, we provide our source code and data at https://github.com/Jyonn/Lamia/ .