Beyond embedding-mapping: Social network alignment via generative information fusion and LLM-guided iterative mechanism

Abstract

Social Network Alignment (SNA) aims to identify and match user accounts belonging to the same real-world individual across multiple social platforms, which has garnered growing research interest. Existing methods typically encode textual and structural information into a latent space and learn a mapping function from annotated user alignments to accomplish SNA. However, the inherent sparsity and noise in social data limit these models’ ability to fully capture user characteristics. Moreover, direct alignment based on latent space often overlooks critical details from the original information, reducing both alignment quality and interpretability. To address these limitations, we propose LLM-SNA, a novel framework that integrates generative information fusion and LLM-guided iterative mechanism. The generative information fusion leverages LLMs to transform sparse user attributes, microblogs, and neighbor descriptions into enriched, comprehensive user profiles. We further perform intra-network and inter-network graph learning on enriched user profiles to incorporate structural information. To balance accuracy and efficiency, the LLM-guided iterative mechanism first applies a coarse filter based on embedding similarities to collect potential alignment candidates. The LLM then evaluates these candidates by reasoning over their original textual information. If the LLM deems the candidates misaligned, the candidate set is expanded until confident matches emerge. Comprehensive experiments on three widely used datasets demonstrate the advantages of LLM-SNA over state-of-the-art baseline methods and highlight the potential of LLMs for SNA tasks.