HiST: Hierarchical Semantic Tree Augmentation for Generative Recommendation

Abstract

Recommendation systems are indispensable for helping users navigate the vast product catalogs of modern online platforms. Recently, generative recommendation has emerged as a promising paradigm due to its unified and flexible formulation, and it has been adopted in various commercial scenarios. However, we find that different tokenization algorithms share a common issue: semantic branch drift, where semantically similar items are often assigned to different branches of the semantic tree. This phenomenon leads to severe information loss in the generated semantic IDs and significantly degrades recommendation accuracy. To address this problem, we propose HiST (Hierarchical Semantic Tree Augmentation), a novel framework that leverages branch information to bridge the semantic gap. The core innovation of HiST lies in using a trained tokenizer to construct a semantic tree over all items and extracting hierarchical information through a bottom-up approach. This hierarchical information is then fused with the original semantic IDs to enhance representation capacity. During end-to-end training, HiST improves recommendation performance by introducing separate prediction objectives for semantic IDs and hierarchical description components, enabling the model to effectively learn both discrete structure and semantic relationships. Experimental results demonstrate that HiST significantly improves recommendation performance across multiple datasets. Furthermore, experiments verify that HiST can be effectively integrated with various existing tokenization methods to enhance recommendation model performance, achieving consistent improvements across evaluated settings.