MANE: A Multi-cascade Adversarial Network Embedding Model for Anchor Link Prediction

Anchor link prediction plays a crucial role in social network analysis as it is a fundamental task. Current studies mainly rely on feature embedding of user-generated content or network structure to create latent representations. These representations are then used to determine user correspondences by mapping users between different network representation spaces. However, most studies still face some challenges such as independent representation spaces, weak discrimination of embedded representations, etc. In this paper, we propose a Multi-cascade Adversarial Network Embedding (MANE) model to tackle these challenges. MANE learns multiple representations from different cascade networks for each user, making users distinguishable in the representation space. Furthermore, an adversarial network is integrated with a mapping function to output a high-quality collection of possible anchor links for correspondence matching. Extensive experiments on real-world social network datasets demonstrate that our method can achieve the expected performance, especially in improving the top-1 precision and recall.