Dynamics of temporal influence in polarised networks

TL;DR

This paper investigates how influence dynamics evolve over time in polarized social networks, proposing methods to track influential users and communities despite changing opinions and network structures.

Contribution

It introduces extended temporal centrality measures that incorporate community structure and evaluates their effectiveness in identifying influential nodes over time.

Findings

Temporal degree centrality and a modified cascade model effectively identify influence bands.

Community-based aggregation reveals influence shifts within polarized networks.

Influence rankings are more stable when accounting for community structure.

Abstract

In social networks, it is often of interest to identify the most influential users who can successfully spread information to others. This is particularly important for marketing (e.g., targeting influencers for a marketing campaign) and to understand the dynamics of information diffusion (e.g., who is the most central user in the spreading of a certain type of information). However, different opinions often split the audience and make the network polarised. In polarised networks, information becomes soiled within communities in the network, and the most influential user within a network might not be the most influential across all communities. Additionally, influential users and their influence may change over time as users may change their opinion or choose to decrease or halt their engagement on the subject. In this work, we aim to study the temporal dynamics of users' influence in a polarised social network. We compare the stability of influence ranking using temporal centrality measures, while extending them to account for community structure across a number of network evolution behaviours. We show that we can successfully aggregate nodes into influence bands, and how to aggregate centrality scores to analyse the influence of communities over time. A modified version of the temporal independent cascade model and the temporal degree centrality perform the best in this setting, as they are able to reliably isolate nodes into their bands.