From Random IDs to Semantic IDs: Building a Generative Recommender from Scratch

2026-05-09 · Source: MLWhiz: Recs|ML|GenAI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, short

Summary

This article introduces Semantic IDs, a novel approach to item representation in recommender systems that addresses the limitations of traditional arbitrary integer IDs. Unlike models like GRU4Rec and SASRec, which treat items as mere numbers, Semantic IDs embed meaningful content information directly into item identifiers. This method, pioneered by the TIGER (Transformer Index for Generative Recommenders) model at NeurIPS 2023, enables recommender systems to generalize to new items, mitigate cold-start problems, and reason about item categories and relationships. The core of this pipeline involves using a Residual Quantized VAE (RQVAE) to compress content-based item embeddings into discrete, token-based Semantic IDs, which then allows a generative model to predict the next item token by token, replacing the traditional retrieve-rank-rerank pipeline.

Key takeaway

For AI Engineers building recommender systems, adopting Semantic IDs can significantly enhance model generalization and address the cold-start problem for new items. By leveraging content-based embeddings and RQVAE, your systems can move beyond memorizing ID transitions to reasoning about item properties, simplifying the recommendation pipeline by replacing separate retrieval and ranking stages with a single generative model.

Key insights

Semantic IDs transform arbitrary item identifiers into meaningful, content-rich tokens for generative recommenders.

Principles

• Content-based IDs improve cold-start.
• Generative models replace retrieve-rank pipelines.
• Progressive compression refines item representation.

Method

The TIGER pipeline uses a text encoder for item embeddings, then Residual Quantized VAE (RQVAE) to generate discrete Semantic IDs, followed by a generative model that predicts next item tokens via beam search.

In practice

• Use RQVAE for item embedding compression.
• Implement beam search for generative recommendations.
• Integrate content embeddings into item IDs.

Topics

• Semantic IDs
• Generative Recommender Systems
• Residual Quantized VAE
• TIGER Recommender
• Cold-start Problem

Best for: Machine Learning Engineer, AI Scientist, AI Engineer

Related on AIssential

• FLUID: From Ephemeral IDs to Multimodal Semantic Codes for Industrial-Scale Livestreaming Recommendation — FLUID replaces ephemeral item IDs in livestreaming recommendation with cross-domain multimodal semantic codes to overcome cold-start limitations.
• Bridging the Semantic-Collaborative Gap: An Asymmetric Graph Architecture for Cold-Start Item Recommendation — An asymmetric graph architecture can bridge the semantic-collaborative gap for cold-start recommendations by aligning intrinsic features with behavioral signals.
• Knowledge Engineering for Search and Content: A Practical Guide — Structured understanding of content and user intent is crucial for effective LLM-powered search and content experiences.
• DiffCold: A Diffusion-based Generative Model for Cold-Start Item Recommendation — DiffCold employs conditional diffusion to unify warm and cold item representations, resolving the "seesaw dilemma" in recommendation systems.
• Structuring and Tokenizing Distributed User Interest Context for Generative Recommendation — G2Rec unifies graph-based co-engagement and semantic tokenization for scalable, accurate generative recommendation.
• Why Generative Recommendation Still Cannot Replace the Full Ranking Pipeline — Generative recommendation exposes the complex, multi-objective nature of industrial recommender systems, hindering end-to-end replacement.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by MLWhiz: Recs|ML|GenAI.