Advancing Semantic ID-based Generative Recommendation

 

Advancing Semantic ID-based Generative Recommendation

 

Speaker: Shutong Qiao  

Host: Dr Ruihong Qiu 

 

Abstract:

Generative recommendation has emerged as a promising paradigm that models recommendation as a sequence generation problem, where items are represented as discrete Semantic IDs (SIDs) and predicted autoregressively. However, existing SID-based frameworks face three key challenges: suboptimal item representation due to the mismatch between real-world attribute-centric metadata and natural language encoders, inefficiency caused by fixed-length identifiers and autoregressive generation, and weak alignment between SIDs and large language models (LLMs).

 

To address the first challenge, we revisit item representation for SID learning and propose an OCR-based paradigm that treats textual content as visual signals. Instead of encoding item descriptions as plain text, we leverage OCR representations to better preserve structural and attribute-level information. This design alleviates issues such as token fragmentation and semantic distortion introduced by conventional text tokenization. Extensive experiments on multiple recommendation datasets demonstrate that OCR-based representations significantly improve Semantic ID quality and consistently enhance recommendation performance. These findings highlight the critical role of representation design in SID-based generative recommendation and provide a more effective approach for modeling real-world item semantics.

 

Bio:
Shutong Qiao is a Ph.D. student at the University of Queensland, Australia. Prior to her candidacy, she received her master's degree from Chongqing University. Her research focuses on generative recommendation, sequential recommendation, and LLM-based recommender systems. She is currently working on improving Semantic ID construction and exploring efficient and scalable generative recommendation frameworks under the supervision of Prof. Hongzhi Yin and A.Prof. Tong Chen.