The Signed Ego Network: Modelling and Analysis Through the Lenses of Online Social Networks

Humans have the ability to communicate at a scale and complexity that is unmatched by any other species on our planet. This capacity has been a key factor in allowing us to develop large-scale societies that are predominant across many diverse areas of the globe. What’s more, since the advent of the internet, individuals and groups can be connected regardless of physical location. This increased connectivity has brought with it many new, emergent phenomena, not all of which are beneficial. Therefore, it is becoming increasingly important to understand the ways in which humans communicate and the social structures that underpin these interactive behaviours. One particularly pertinent model that facilitates this understanding is the Ego Network Model (ENM), which views a social network from the point-of-view of a single individual, organising their connections into circles around them based on the strength of their relationship. However, this model has the notable limitation of only measuring relationships based on the contact frequency of the individuals involved. This thesis aims to establish an extension to the ENM that incorporates signed connections: the Signed Ego Network Model (SENM). To this end, a novel methodology for computing polarity sign (i.e. positive or negative) for the connections within a social network, based on sentiments of individual interactions, is first proposed. This method is shown to achieve similar results regardless of the model used to compute the individual sentiments and is also validated using known expectations of signed networks. Next, the signing methodology is used to compute the SENM, which is then investigated. The results of this reveal that, surprisingly, negativity is often most prevalent in the relationships we engage the most with. The potential effects of negativity on cognitive load are also investigated, although little statistically significant evidence was found. The ENM and SENM are then leveraged for the task of Stance Detection (SD), where they are able to be used to obtain results similar (although slightly worse) to the cutting edge, while using far less, and more easily obtainable, data. Finally, differences in the SENM are observed between cultures and online communities. Both cultures and engagement in a subcommunity were found to have an effect on the rate of negative relationships, although the former appears to be more influential. This is followed up by analyses of the most popular terms and talking points, which find that individuals who engage in more generic exchanges (e.g. about the weather) are more likely to have fewer negative relationships than those who commonly engage in more polarising topics (such as politics).