UC San Diego

The rapid development of large generative models has motivated a shift in recommender systems toward deep generative architectures with compact vocabularies. A fundamental operation in this paradigm is action tokenization: the process of converting human-perceivable actions, such as user-item interactions, into machine-readable token sequences, often referred to as semantic IDs. In this dissertation, we investigate the role of action tokenization in generative recommendation, focusing on how human actions should be tokenized and leveraged. Our research follows three main directions. First, we develop stronger action tokenization methods that make semantic IDs more expressive and effective, including scaling semantic IDs from short to long sequences and moving from context-independent to context-aware tokenization. Second, we study how to better leverage semantic IDs during training and inference so that generative models can use them more effectively, including improving new-item recommendation and addressing the personalization limits introduced by autoregressive token generation. Third, we show that generative recommendation models can excel at generalization, and empirically demonstrate that this capability is closely tied to their action tokenization mechanisms. Together, these findings advance both the methodology and understanding of action tokenization in generative recommendation, providing a foundation for developing future recommender systems powered by large generative models.