Semantic IDs for Generative Search and Recommendation

Researchers at Spotify developed a multi-task approach for creating Semantic IDs that work effectively in unified generative models for both search and recommendation. Traditional task-specific Semantic IDs fail to generalize across both functions, but their proposed bi-encoder model jointly fine-tuned for both tasks achieves balanced performance. The multi-task Semantic IDs represent items as discrete tokens based on embeddings that incorporate information from both search queries and recommendation patterns, offering a promising direction for more generalizable generative recommender systems.