G2Rec: Tokenizing User Interest Graphs for Scalable 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 graph-based integration methods, such as graph serialization and graph neural networks, either suffer from scalability issues or exploit only local graph information. Similarly, existing semantic tokenization methods typically rely on heuristics and lack explicit supervision signals, which may lead to inaccurate or suboptimal semantic representations.

To address these limitations, 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 user-item interaction graphs and semantic item tokens, G2Rec enables recommendation models to capture holistic and semantically grounded user interest prototypes without requiring ground-truth user interests. This provides more comprehensive and accurate modeling of user behavior contexts in industrial sequential recommendation. The framework has been deployed online across product surfaces and validated through extensive experiments on public datasets, demonstrating superiority over existing methods.

For the crypto and blockchain domain, G2Rec’s approach offers several concrete implications. First, the generative tokenization of user-item interactions can be directly applied to on-chain transaction sequences. By treating each transaction or smart contract call as a ‘token’ in a generative model, one can predict the next state transition (e.g., next swap, liquidation, or cross-chain message) from historical on-chain data. This could replace heuristic mempool strategies with learned generative models for MEV searchers or intent solvers.

Second, G2Rec’s graph-based co-engagement modeling can scale to millions of wallet addresses without ground-truth labels. DeFi protocols could construct co-engagement graphs of wallet addresses (who transacts with whom, which pools they share) to tokenize ‘user interest contexts’ for personalized liquidations, yield strategies, or credit scoring. This unsupervised approach is particularly valuable in permissionless blockchains where users never explicitly declare their intent.

Third, the semantic tokenization of items into a unified vocabulary can be extended to tokenize smart contract interactions, NFT metadata, or token symbols. This enables a new class of ‘tokenized blockchain state’ models that predict next interactions as a generative language modeling task, with applications in fraud detection, wallet UX, and automated DeFi agents.

Finally, G2Rec’s proven online deployment at industrial scale suggests that similar architectures could be deployed as on-chain or off-chain inference pipelines for a ‘recommendation layer’ in DeFi—e.g., suggesting optimal swap routes, yield pools, or NFT bids based on a wallet’s tokenized on-chain history, with the scalability to handle Ethereum’s full historical transaction volume.