The emergence of large generative models is transforming the landscape of recommender systems, offering new opportunities through scaling laws and flexible content generation. One of the most fundamental components is action tokenization, which is the process of converting human-readable data (e.g., text or user-item interactions) into model-readable token sequences. This talk provides a tokenization-centric view of building effective and efficient generative recommendation models. We begin by introducing the backgrounds of semantic IDs and action tokenization techniques, as covered in our WWW’25 tutorial. We then introduce two recent work: aligning action tokens with LLMs (ICDE’24) and contextually tokenizing action sequences (ICML’25 spotlight).
Speaker Bio
Yupeng Hou is a Ph.D. student at the University of California, San Diego, advised by Prof. Julian McAuley. He was a student researcher at Google DeepMind in 2024, working with Jianmo Ni, Derek Cheng, and Ed H. Chi. He previously received his M.E. and B.E. from Renmin University of China, advised by Prof. Wayne Xin Zhao. His work has been recognized as the Best Resource Paper Runner-up at CIKM 2022 and the Best Student Paper Runner-up at RecSys 2022. Yupeng is one of the leading developers of RecBole, a popular open-source recommendation library with over 3,700 GitHub stars. His current research focuses on generative recommendation and tokenization.
More Details
- When: Thursday. 22 May 2025, at 1-2pm (Brisbane time)
- Speaker: Yupeng Hou (UCSD)
- Host: Dr Ruihong Qiu
- Zoom: https://uqz.zoom.us/j/86490391734
Would prompt work for graph learning? An exploration of few-shot learning on graphs