Recommending Smart Contracts to Users Using Semantic Graph Embeddings

Abstract

Public blockchains like Ethereum generate vast amounts of transactional data, offering insight into significant economic and decentralized application activity. However, these data are often unstructured and semantically poor. Understanding the functionality of smart contracts is crucial for unlocking the potential of this data. This work presents a novel methodology to transform raw blockchain transaction logs into semantically rich representations. We first classify smart contracts by analyzing their function names and structures using N-gram profiling and embedding comparisons against known standards like ERC/EIP. This process assigns functional labels (e.g., DeFi, Exchange, Token) to the smart contracts. Subsequently, we model the framework as a heterogeneous graph of users and labeled contracts. We employ a modified Node2Vec algorithm with an enforced alternating node-type walk strategy to effectively capture the dynamics of user-contract interactions. This process yields low-dimensional vector embeddings for users and contracts, making blockchain data readily available for knowledge discovery. We demonstrate the utility of our approach through a smart contract recommendation system that suggests relevant contracts to users based on their interaction history and learned embeddings.