An approach based on multiplex networks for modeling cascading trust failures in social networks

摘要

Removing nodes or links from a real-world social network may lead to a collapse in the entire network itself. This is due to the propagation effect of the initial removal. In the literature, this phenomenon is called cascading failure. In the context of trust modeling, cascading failure occurs when a node's trust toward another, changes to distrust resulting in the removal of the trust link between them. This change in the trust network may affect the other nodes' trust toward the target node. As the number of failures in a network increases, the users become more reluctant to share their interests and opinions with other members. Hence, it is important to model, detect and mitigate the cascading trust failures. Currently, simple computational trust models are used for modeling cascading trust failures in the existing works. The effect of relevant contexts in modeling cascading trust failures is still a missing concept in the proposed models. Failure in a specific trust context may impact the relevant contexts as well and may have less or no impact on the irrelevant contexts. Therefore, the need for a more comprehensive trust modeling approach for cascading trust failures is evident. In this paper, the proposed computational trust is formulated by considering the context dependencies in addition to the impact of trust contexts on one another. Also, by mapping the trust contexts to a multiplex network's layers and using the advantages of complex network analysis concepts, a new method for computing the similarity between the trust contexts is introduced. The introduced trust model uses the trust information of all layers (i.e., contexts) to compute the new trust values after a trust failure. In addition, the trust model uses the newly provided information to adjust the calculated trust values by leveraging real-world data. Besides, a model for trust cascading failure as well as an attack prevention method are introduced. By performing a wide range of experiments, including sensitivity analysis, accuracy analysis, and comparative studies, the effectiveness of the proposed approach is evaluated. Real-world networks' data such as Facebook and Twitter's ego nets and synthetic data are used for the performed evaluations. It is shown that higher values for the context's importance parameters make the trust links more vulnerable and easier to fail. The three well-known trust attack scenarios including HT, LT, and RT are performed and it is demonstrated that the layers with high similarity values tend to have more similar cascading failure patterns. It is shown that the accuracy of the introduced model for trust cascading failures is higher compared to the existing works. By adding the attack prevention component, the model's accuracy gets close to 1 in best cases, which is a notable improvement.