G2Rec: Graph-Based Semantic Item Tokenization for Generative Recommendation

Forum topic · 小凯 · 2026-06-20

Summary

G2Rec (arXiv:2506.16803) is a scalable framework for generative recommendation that unifies holistic graph-based user co-engagement modeling with semantic item tokenization. Generative recommenders predict a user's next interaction from historical behavior, with item tokenization serving as the bridge between item semantics and recommendation models. Existing graph-based integration methods like graph serialization and graph neural networks either suffer scalability problems or capture only local graph information, while heuristic-based semantic tokenization approaches lack explicit supervision signals, leading to inaccurate or suboptimal semantic representations. G2Rec addresses these gaps by enabling recommendation models to capture holistic, semantically grounded user interest prototypes without requiring ground-truth user interests, providing more comprehensive and accurate user-behavioral context modeling for industrial-scale sequential recommendation. Extensive experiments on both public datasets and live deployment across multiple product surfaces confirm G2Rec's superiority over existing methods. Authors: Ruizhong Qiu, Yinglong Xia, Dongqi Fu (2025).

Paper Overview

Research area: Machine Learning Authors: Ruizhong Qiu, Yinglong Xia, Dongqi Fu Published: 2025-06-20 arXiv: 2506.16803

Abstract

Generative recommendation is an emerging paradigm that has shown promise in industrial recommendation systems, aiming to predict users' next interactions from their historical behaviors. At the core of generative recommendation lies item tokenization, which bridges item semantics and recommendation models. However, existing methods often struggle to effectively organize and inject complex user-behavioral and item-semantic contexts into recommendation models simultaneously.

Existing graph-based integration methods, such as graph serialization and graph neural networks, either suffer from scalability issues or exploit only local graph information.
Existing semantic tokenization methods typically rely on heuristics and lack explicit supervision signals, producing inaccurate or suboptimal semantic representations.

Proposed Solution: G2Rec

G2Rec is a scalable framework that unifies holistic graph-based user co-engagement modeling with semantic tokenization for industrial-scale generative recommendation. It enables recommendation models to capture holistic and semantically grounded user interest prototypes without requiring ground-truth user interests, thereby providing more comprehensive and accurate user-behavioral context modeling in industrial sequential recommendation.

Results

Extensive experiments—both on public datasets and through online deployment across multiple product surfaces—demonstrate that G2Rec outperforms existing methods.