Enhancing trust prediction in attributed social networks with self-supervised learning

Predicting trust in Online Social Networks (OSNs) is essential for a range of applications including online marketing and decision-making. Traditional methods, while effective in some scenarios, encounter difficulties when attempting to handle the complexities of trust networks and the sparsity of trust relationships. Current techniques attempt to use user attributes such as ratings and reviews to fill these data gaps, although this approach can introduce noise and compromise prediction accuracy. A significant problem remains: most users do not explicitly state their trust relationships, making it difficult to infer trust from a vast amount of unlabelled data. This paper introduces a novel model, Trust Network Prediction (TNP), which employs self-supervised learning to address these issues within attributed trust networks. TNP learns efficiently from unlabelled data, enabling the inference of potential trust connections even without explicit trust relationships. It also minimises redundancy and the impact of abundant unlabelled data by generating comprehensive user representations based on existing trust relationships and reviewing behaviour. An advanced model, the Time Series Trust Network Prediction (TSTNP), extends TNP’s functionality by considering the temporal nature of trust in OSNs. By incorporating time series data, TSTNP analyses how user review behaviour at different time points influences trust prediction. Through comprehensive testing on two real-world datasets, our proposed models demonstrate their effectiveness and reliability in trust prediction tasks, underscoring their potential utility in OSNs.