AsyLink: user identity linkage from text to geo-location via sparse labeled data

Abstract

User Identity Linkage (UIL) aims to reveal the correspondence among account pairs across different social platforms. It has been a popular but challenging task in recent years as complex application scenarios have emerged. Existing UIL methods mainly formalize a classification problem based on symmetric information, but these techniques are hard to apply to asymmetric, sparsely labeled, and imbalanced data. To combat the challenges, we propose a novel UIL framework (AsyLink) with asymmetric information in text and geographic forms. AsyLink first uses topic modeling technologies to associate words and locations, where external text-location pairs can be conveniently introduced to reduce bias caused by sparse linkage labels. Then the user-user interactive tensors are constructed as the basis for linking. Using 3D convolutional neural networks, matching patterns in user-user interactive tensors are captured, and final predictions are based on the extracted features. Meanwhile, instead of regular classification loss, the ranking loss is introduced to predict the best answer among candidates, which is conducive to imbalanced classification. Experiments performed on four real-world datasets indicate that AsyLink achieves state-of-the-art performances and has great potential for real-world applications.