Abstract
This thesis is concerned with predicting social influence on social networks and online social networks in particular. We consider online social networks consisting of users on a social media platform and their connections on the platform. Social influence refers to the phenomenon that opinions, ideas and behaviours of individuals are affected by other individuals with whom they are connected via social connections. Analysing social influence is useful in many applications, such as virtual marketing, recommendation, online advertising and expert detection. Social influence predicting is an important area of social influence analysis.Some social influence analysis studies have focussed on how social influence diffuses on social networks and how to maximize social influence. This thesis focuses on the prediction of social influence on individuals and studies two research problems, namely, which users produce more influence on social networks and which users are influenced by others.
Current research on the problem of finding users who produce more influence is focused on proposing influential user detection algorithms. These detection algorithms commonly use the global structure of social networks and the complex relationships between users to model influence spreading. However, different types of interactions between users are overlooked in these algorithms. This thesis considers if interaction is an important factor in producing influence and if different types of interactions need to be treated differently. Based on this, it proposes a novel influential user detection algorithm based on interactions between users.
In the problem of predicting users who are influenced by others, current research tries to predict user behaviour change under others' influence by using deep learning approaches. Due to the complexity of social networks, several gaps have not been considered in the current work. Firstly, the datasets collected from social media may have quality issues such as missing connections between users. Secondly, for emerging topics, there is a lack of available labels. Thirdly, social networks are constantly changing when new topics appear, and it is not cost-effective to retrain the model for different topics. This thesis proposes new methods to overcome the above gaps with state-of-the-art graph machine-learning techniques.
Four main chapters in this thesis present the proposed new approaches to bridging the four research gaps. All approaches are evaluated against real social network datasets collected from well-known social media platforms. The evaluation results of the proposed approaches in the thesis outperform the existing state-of-the-art methods. Finally, the thesis discusses the limitations of proposed approaches that can be further improved by researchers in the field.
Date of Award | 9 May 2023 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | Weiru Liu (Supervisor), Ryan McConville (Supervisor) & Jun Hong (Supervisor) |