Influence-based detection and mitigation of ideological isolation in online social networks

Ideological isolation such as the echo chamber effect and the filter bubble effect, are criticised for restricting individuals in their prior beliefs and excluding them from diverse viewpoints. The quick development of Online Social Networks (OSNs), along with their employed Artificial Intelligence (AI) recommendation techniques has made ideological isolation become an unavoidable concern. The detection and mitigation of ideological isolation help to improve these effects and their negative consequences (e.g. network polarisation and the spread of misinformation) and make OSNs better function as public spheres.
In this thesis, the detection and mitigation of ideological isolation are studied step-by-step. To begin with, a context-aware influence diffusion model (CAID) is proposed. CAID considers individuals’ contexts, which are neglected by existing methods for modelling ideological isolation, in the influence diffusion process. Two scenarios based on a real-world dataset are conducted to track the pattern of the proposed CAID model. The simulation results proved the selective exposure mechanism behind ideological isolation, the higher belief an individual holds, the more likely for his/her to adopt information that favours his/her prior beliefs. This result suggests that CAID is applicable for modelling ideological isolation in OSNs. After that, a context-based modelling approach, the Recommendation-based Influence Diffusion Model (RIDM) is developed to model ideological isolation in OSNs with recommendation services. Two quantification metrics, i.e., the echo chamber and the filter bubble indices, are designed to measure these effects separately under the proposed RIDM. Several simulations are conducted on two real-world datasets to explore the impact of AI recommendation on the evolution of ideological isolation, in the forms of the echo chamber and filter bubble effects. Simulation results suggest recommendation algorithms facilitate the filter bubble effects, leading users to become ideological isolation at an individual level. Whereas, at a topological level, recommendation algorithms show their ability to connect dissimilar users or recommend diverse viewpoints, tempering the echo chamber effects. In the end, a Nudge-based Adaptive Recommendation Strategy (NARS) is introduced to mitigate the ideological isolation in OSNs. Instead of simply refuting prior beliefs with opposite viewpoints, NARS gradually and gently directs users to step out of ideologically isolated environments via mild nudges. Notably, NARS inherits the advantages of the Memetic Algorithm and therefore captures the dynamics of users’ beliefs in a changing topic environment, where the existing topics keep changing. By leveraging a specific selection operator, NARS is adaptive to users’ feedback. With the comparison to several baseline recommendation methods, NARS shows the best performance in mitigating ideological isolation.
The contribution of this project is summarised as:
• A novel Context-Aware Influence Diffusion model CAID to model the ideology of an individual in OSNs.
• A novel Recommendation-based Influence Diffusion Model RIDM to model ideological isolation in OSNs with recommendation service.
• Two quantification metrics, the echo chamber indices and the filter bubble indices, to measure these two effects separately in online social networks, and to explore the impact of Recommendation algorithms on causing ideological isolation.
• A Nudge-based Adaptive Recommendation Strategy, NARS to mitigate the ideological isolation via mild nudges.