RASTP: Representation-Aware Semantic Token Pruning for Generative Recommendation with Semantic Identifiers
Tianyu Zhan, Kairui Fu, Zheqi Lv, Shengyu Zhang
TL;DR
This work tackles the computational overhead of SID-based generative recommendations by introducing Representation-Aware Semantic Token Pruning (RASTP). RASTP selects the most informative semantic tokens by combining semantic saliency, captured by token representations, with attention centrality from the model's internal attention, retaining only a subset of tokens during training. Experiments on three real-world Amazon datasets demonstrate a 26.7% reduction in training time without sacrificing performance, and ablation studies show that pruning timing and strategy significantly influence efficiency and accuracy. The approach offers a practical path to more scalable SID-based generation in industrial settings and is released as open-source code for replication and extension.
Abstract
Generative recommendation systems typically leverage Semantic Identifiers (SIDs), which represent each item as a sequence of tokens that encode semantic information. However, representing item ID with multiple SIDs significantly increases input sequence length, which is a major determinant of computational complexity and memory consumption. While existing efforts primarily focus on optimizing attention computation and KV cache, we propose RASTP (Representation-Aware Semantic Token Pruning), which directly prunes less informative tokens in the input sequence. Specifically, RASTP evaluates token importance by combining semantic saliency, measured via representation magnitude, and attention centrality, derived from cumulative attention weights. Since RASTP dynamically prunes low-information or irrelevant semantic tokens, experiments on three real-world Amazon datasets show that RASTP reduces training time by 26.7\%, while maintaining or slightly improving recommendation performance. The code has been open-sourced at https://github.com/Yuzt-zju/RASTP.