IDs That Mean Something: A Hands-On Tutorial on Semantic IDs for Recommender Systems
Overview
Recommender systems have undergone a significant transformation in how they represent and identify items, evolving from raw IDs as simple lookup keys to semantic IDs as structured interfaces for utilizing and sharing information across recommendation settings. This tutorial reviews that evolution — raw IDs, auxiliary content, and semantic IDs — and their applications across multimodality, cold-start, and cross-domain recommendation, complemented by two hands-on sessions.
Learning Goals
By the end of this tutorial, attendees will:
- Have the conceptual grounding and practical tools to apply semantic IDs in their own research or production systems.
- Be able to identify scenarios where semantic IDs have the most impact, including cold-start and cross-domain recommendation.
- Understand open challenges and emerging directions beyond conventional recommendation, such as dynamic semantic IDs, multi-stakeholder utility, and semantic planning.
Target Audience
This tutorial targets researchers and practitioners with a general interest in recommender systems, information retrieval, and representation learning — particularly those working on knowledge representation and scalable machine learning.
Prerequisites:
- Basic familiarity with machine learning concepts and recommender systems fundamentals.
- No prior knowledge of semantic IDs or generative retrieval is required.
Tutorial Outline and Structure
The half-day tutorial consists of 3 hours of core content, plus two coffee breaks. All start times are tentative and will be finalized once CIKM '26 releases the official program.
| Time (tentative) | Section | Presenter |
|---|---|---|
| 9:00–9:15 | Introduction to recommender systems, raw item IDs, and auxiliary content | |
| 9:15–9:30 | Semantic IDs: introduction & overview | |
| 9:30–9:50 | Leveraging semantic IDs in recommender systems | |
| 9:50–10:05 | Coffee break | |
| 10:05–10:30 | Hands-on: building and using semantic IDs | |
| 10:30–10:55 | Multimodality and collaborative signals | |
| 10:55–11:20 | Cold-start mitigation and cross-domain transfer | |
| 11:20–11:35 | Coffee break | |
| 11:35–12:00 | Hands-on: cross-domain usage of semantic IDs | |
| 12:00–12:15 | Future directions and open challenges | |
| 12:15–12:30 | Q&A and open discussion |
ℹ️ Exact session date, start time, and per-section presenter assignments will be announced once CIKM '26 releases the official program.
Format: both hands-on sessions use the Amazon Reviews dataset, with instructions and code provided as self-contained Jupyter notebooks that run on free-tier resources such as Google Colab. Content-based item embeddings and baseline models are pre-computed so the sessions stay focused on semantic IDs. Attendees should bring their own laptop for the hands-on sessions; presentation requires an LCD projector, at least one projection screen, and microphones.
Presenters
Roger Zhe Li
Huawei Ireland Research Center
Aghiles Salah
Huawei Ireland Research Center
Mete Sertkan
Huawei Ireland Research Center
Changhong Jin
University College Dublin
Yingjie Niu
Huawei Ireland Research Center
Ruihai Dong
University College Dublin
Barry Smyth
University College Dublin
Citation
If you find this tutorial useful, please consider citing:
@inproceedings{li2026idsmeansomething,
title = {IDs That Mean Something: A Hands-On Tutorial on Semantic IDs for
Recommender Systems},
author = {Li, Roger Zhe and Salah, Aghiles and Sertkan, Mete and Jin, Changhong
and Niu, Yingjie and Dong, Ruihai and Smyth, Barry},
booktitle = {Proceedings of the 35th ACM International Conference on Information
and Knowledge Management (CIKM '26)},
year = {2026},
isbn = {979-8-4007-2539-5},
location = {Rome, Italy},
numpages = {5},
doi = {10.1145/3799682.3839274},
url = {https://doi.org/10.1145/3799682.3839274}
}Quick Links
- Papers
- Code & Notebooks — to be released upon acceptance
Maintained by Roger Zhe Li — Last updated: 2026