Deep trust prediction in online social networks

Online Social Networks (OSNs) have become indispensable in our daily life. While the popularity of social networking consistently rises, it is inevitable to raise new issues for online users, such as information overload and information credibility. To address these issues, trust prediction is critical to provide an effective way for online users to decide with whom to share information and from whom to accept information. Traditional trust prediction works usually suffer severely from the data sparsity problem. In addition, the various factors, especially the latent factors, which may affect trust prediction performance have not been well considered.

In this thesis, we aim to develop effective trust prediction approaches based on deep learning techniques to fill the aforementioned gaps. To alleviate the long-standing data sparsity problem suffered in traditional trust prediction methods, we first propose a novel data sparsity insensitive deep user model for trust prediction based on the social homophily theory. Then, we propose a context-aware deep trust prediction model by considering both the users’ static preference features and dynamic preference features. In addition, network embedding has achieved promising performance for link prediction by learning node representations that encode intrinsic network structures. Accordingly, we develop a network embedding-based method that incorporates both trust properties and trust network structures. Finally, we propose a novel multi-Graph Attention Network-based approach to apply the obtained trust relationships on the recommendation tasks. Extensive experiments have been conducted against real-world datasets to demonstrate the effectiveness of the proposed models.