MFLink: User Identity Linkage Across Online Social Networks via Multimodal Fusion and Adversarial Learning

As an essential step in the online social network research, user identity linkage aims to identify different accounts belonging to the same natural person. Many existing methods rely on single-modal approaches, which cannot provide a comprehensive user description. Some methods also fail to address the semantic gaps in data across different social platforms. To concurrently address these issues, this paper explores user identity linkage across online social networks by leveraging three types of modal information of users: attributes, post content, and social relationships. We propose a user identity linkage scheme named MFLink based on multimodal fusion, which has three components: Feature Extraction, Multimodal Fusion, and Adversarial Learning. In the Feature Extraction, MFLink utilizes feature embedding methods to transfer the user attribute and post content into intermediate representations. To achieve optimal fusion of information from these three modalities, MFLink integrates each modality with the assistance of graph neural networks and an attention mechanism within the Multimodal Fusion. Finally, MFLink employs adversarial learning to enhance the similarity of representations for the same individual across various platforms. The experiment results on the TWFQ dataset indicate that MFLink outperforms the advanced approaches in fusing information of modalities and addressing the data semantic gaps across online social networks.