Using Agent-Based Modelling to Investigate Intervention Algorithms to Reduce Polarisation in Online Social Networks

Coates, Adam (2020) Using Agent-Based Modelling to Investigate Intervention Algorithms to Reduce Polarisation in Online Social Networks. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (9MB) | Preview

Abstract

Across much of the western world, political polarisation is on the rise. This has the effect of hindering political discourse, stifling open discussion, and in extreme cases has led to violence. The process of polarising and radicalising vulnerable individuals has migrated to social media websites, which have been implicated in several high profile terror attacks. Within this thesis we model and investigate various algorithms to prevent the spread of polarisation and extremist ideology by employing agent-based modelling techniques from the field of opinion dynamics. The contributions of our work include the following aspects. Firstly, we have developed a unified framework for opinion dynamics, allowing us to experiment easily on a number of different existing models and bringing together sometimes disparate innovations from across the field into one system. Secondly, this unified framework has been implemented in a modular simulator able to perfectly replicate results from purpose-built, stand-alone simulators for two widely used models, namely Relative Agreement and CODA, and then released to the public as the first general-purpose opinion dynamics simulator. Thirdly, we have developed two new intervention algorithms, along with a new metric for measuring the effectiveness of an intervention strategy, which aim to reduce the spread of polarisation across a network with low computational cost. These methods are compared to existing centrality-based methods upon a random network. The experimental results show our proposed approaches outperform centrality measures. We find that our ii iii algorithms are able to prevent up to 40% of non-extremist agents becoming extreme by removing only 10% of the network’s edges. Fourthly, we have investigated the efficacy of these intervention algorithms on polarisation under different scenarios (e.g. variable costs, different network structures). The experimental validation proves the proposed approach is robust and has performed favourably compared existing methods such as centrality-based methods especially on the second type of network. Finally, we have developed a broadcast-based communication system for agents, designed to mimic the one-way broadcast nature of a public social media post such as Twitter, in contrast to the existing model which emulates a two-way private conversation. The experimental result shows a lessening of the impact of our interventions, demonstrating the need for further investigation of such communication methods.