A Diffusion-Based Triple Embedding Model for User Identity Linkage Across Social Networks

User identity linkage identifies anchor users with multiple accounts across different social networks. Current approaches primarily use embedding techniques to match users’ content and structural features but struggle with challenges such as similarity dilemmas, network structure variations, and directional linkages. To address these issues, we propose a diffusion-based triple embedding framework (DTE) that leverages information dissemination behaviors during cross-network information diffusion. Each information diffusion is modeled as a triplet within a diffusion network, transforming the linkage problem into a cross-network triplet prediction task. We extract two types of triplet contexts and incorporate them into the learning process, supervised by a small set of anchor user pairs. Our designed triple translation in complex space ensures linkage by maintaining the equivalence between anchor accounts. Experimental results demonstrate that DTE outperforms state-of-the-art methods.