IDs That Mean Something

A hands-on CIKM '26 tutorial on semantic IDs for recommender systems — from raw integer IDs to semantic IDs, across generative and discriminative applications.

Overview

Traditional recommender systems rely on arbitrary integer IDs to identify items — identifiers that carry no semantic information and force models to learn representations purely from interaction data. Semantic IDs address this by encoding semantic information directly into the item identity itself: items become discrete token sequences that reflect content, hierarchy, and relationships, rather than opaque lookup keys.

This tutorial traces the evolution from raw IDs to semantic IDs as a scalable means of information utilization in recommender systems, offering a comprehensive review of their development and applications across multimodality, cold-start, and cross-domain recommendation. Hands-on exercises guide attendees through building and evaluating semantic IDs, training models, and exploring cross-domain transfer, before closing with open challenges such as dynamic semantic IDs and multi-stakeholder utility.

This tutorial covers:

The evolution from raw IDs → auxiliary content → semantic IDs, and why each stage falls short.
Semantic IDs in both generative (e.g., TIGER) and discriminative (e.g., YouTube ranking) recommendation.
Multimodality, collaborative signals, and cross-domain transfer via semantic IDs.
Hands-on construction and evaluation of semantic IDs on the Amazon Reviews dataset.

This site curates:

Tutorial materials — session outline, learning goals, and hands-on notebooks.
Papers — curated bibliography spanning raw IDs to semantic IDs.
Presenters — the seven researchers behind the tutorial.