Social influence and influence diffusion on Twitter
With the recent surge of online social networks (OSN), knowledge on how influence spreads across digital spheres has become increasingly valuable in many applications including decision making, viral marketing, event detection and rumour control. Crucial challenges in this area include the difficulty in measuring social influence and its diffusion. Further, the underlying mechanisms in influence diffusion are often complex and likely specific to the social network data. As such, this thesis studies user-to-user influence and the mechanisms through which influence diffuses in OSNs. Specifically, it aims to address three problems relative to the chosen research platform, Twitter: (1) identify and measure user-to-user influence; (2) model influence diffusion for a particular user's action; and (3) model the diffusion of topics throughout a network. First, user-to-user influence on Twitter is examined based on both action and interaction history. The user-to-user influence is measured by a given user's reactions to another user's online activity. To learn the influence probabilities, we utilise both the explicit reactions that are currently available on the platform, including "like" (which is often ignored in existing works), and study implicit reactions by considering following relations, time and content similarity. We show the quality of the learned user-to-user influence probabilities on two tasks -influence diffusion prediction and influential nodes identification – by comparing our proposed methods with a number of baselines. Second, a time- and depth-sensitive diffusion model is proposed to predict the cascade of a particular user's action. The dynamics of influence decay with time and depth when going through the network are analysed in detail and incorporated into the proposed model. Experimental results show that the size, shape, duration and involved users of an influence-diffusion cascade can be predicted. For this, the proposed model outperforms existing ones in terms of balanced precision and recall, especially in predicting when people will be influenced and the shape that a cascade takes. Third, topic diffusion cascades throughout a network are investigated, with particular attention to the formation of bursts. Effective methods are developed to characterise user topic dynamics and to detect topic diffusion bursts. Considering that a user's decision to participate in a topic diffusion cascade is determined by one's own interests and observation of others' actions, a topic-aware conformity diffusion model is proposed. The model covers the behaviour of both source participants and conforming participants. It also captures multiple instances of one's participation in one cascade and the characteristics of possibly diminishing or increasing marginal returns of an additional infected neighbour. Upon evaluating the proposed model with several datasets, results show its ability to reproduce bursts in a given topic diffusion. Altogether, this work provides fundamental insights into influence diffusion in OSNs, including the characteristics of user-to-user influence relationships, the dynamics of influence decay with time and depth, the growth and drop of diffusion cascades for different topics, and the topic-wise conformity of users.