G2Rec Unifies Graph Modeling and Semantic Tokenization for Generative Recommendation

Generative recommendation is an emerging paradigm that aims to predict users’ next interactions from their historical behaviors, with item tokenization at its core bridging item semantics and recommendation models. However, existing approaches face two key limitations: graph-based integration methods (graph serialization and graph neural networks) either suffer from scalability issues or exploit only local graph information, while semantic tokenization methods typically rely on heuristics and lack explicit supervision signals, leading to inaccurate or suboptimal semantic representations.

To address these challenges, researchers propose G2Rec, a scalable framework that unifies holistic graph-based user co-engagement modeling with semantic tokenization for industrial-scale generative recommendation. By jointly learning from the full user-item interaction graph and semantic embeddings, G2Rec enables recommendation models to capture holistic and semantically grounded user interest prototypes without requiring ground-truth user interests. This unsupervised approach is particularly valuable in environments where explicit preference labels are unavailable.

The framework has been deployed online across product surfaces and evaluated on public datasets, demonstrating superiority over existing methods. According to the paper, G2Rec provides more comprehensive and accurate modeling of user behavior contexts in industrial sequential recommendation. Its ability to scale to large graphs while capturing global user co-engagement patterns marks a significant advance over prior graph-based methods that were limited to local neighborhoods or could not handle industrial-scale data.

For the crypto and blockchain domain, G2Rec’s design offers direct parallels. On-chain user-item interaction graphs—such as DeFi protocol usage, NFT trading networks, or wallet-to-contract calls—are massive, sparse, and lack ground-truth user interests. G2Rec’s holistic graph modeling could be repurposed to build scalable, globally-aware recommendation for dApp discovery or token swaps, where local neighborhood methods miss cross-protocol interest patterns. The learned tokenization from user co-engagement could produce semantically meaningful “intent tokens” for predicting the next on-chain action in a wallet’s sequence, replacing heuristic methods like raw ABI encoding. Moreover, G2Rec’s unsupervised interest discovery could cluster wallets by behavioral prototypes (e.g., “liquidity provider,” “NFT flipper”) purely from on-chain data, enabling privacy-preserving personalization without KYC or preference surveys. The fact that G2Rec has been deployed at industrial scale suggests its architecture is production-ready for high-throughput on-chain data streams, such as mempool transaction sequences or cross-chain swap routing.