Recent advancements in generative recommenders have revealed challenges in the two-stage paradigm of tokenization and user interaction modeling. The study identifies key limitations such as suboptimal static tokenization that does not account for the diversity of user contexts, and the problem of overwriting pretrained semantic knowledge during training. Addressing these issues, the authors propose DEcomposed COntextual Token Representations (DECOR), which harmonizes semantic preservation with improved adaptability in token embeddings.
DECOR leverages contextualized token composition to refine embeddings based on user interaction contexts while incorporating decomposed embedding fusion that combines pretrained knowledge with newly acquired collaborative embeddings. This approach not only enhances recommendation accuracy but also addresses the misalignment of objectives in existing frameworks. Extensive experiments on three real-world datasets validate DECOR’s superior performance, suggesting implications for future recommender systems focusing on adaptability and semantic retention.
👉 Pročitaj original: arXiv AI Papers
