Dynamics of Information Diffusion

Abstract

Real diffusion networks are complex and dynamic, since underlying social structures are not only far-reaching beyond a single homogeneous system but also frequently changing with the context of diffusion. Thus, studying topic-related diffusion across multiple social systems is important for a better understanding of such realistic situations. Accordingly, this thesis focuses on uncovering topic-related diffusion dynamics across heterogeneous social networks in both model-driven and model-free ways. We first conduct empirical studies for analyzing diffusion phenomena in real world systems, such as new diffusion in social media and knowledge transfer in academic publications. We observe that large diffusion is more likely attributed to interactions between heterogeneous social networks as if they were in the same networks. Thus, external influences from out-of-the-network sources, as observed in previous work, need to be explained with the context of interactions between heterogeneous social networks. This observation motivates our new conceptual framework for cross-population diffusion, which extends the traditional diffusion mechanism to a more flexible and general one. Second, we propose both model-driven and model-free approaches to estimate global trends of information diffusion. Based on our conceptual framework, we propose a model-driven approach which allows internal influence to reach heterogeneous populations in a probabilistic way. This approach extends a simple and robust mass action diffusion model by incorporating the structural connectivity and heterogeneity of real-world networks. We then propose a model-free approach using informationtheoretic measures with the consideration of both time-delay and memory effects on diffusion. In contrast to the model-driven approach, this model-free approach does not require any assumptions on dynamic social interactions in the real world, providing the benefits of quantifying nonlinear dynamics of complex systems. Finally, we compare our model-driven and model-free approaches in accordance with different context of diffusion. This helps us to obtain a more comprehensive understanding of topic-related diffusion patterns. Both approaches provide a coherent macroscopic view of global diffusion in terms of the strength and directionality of influences among heterogeneous social networks. We find that the two approaches provide similar results but with different perspectives, which in conjunction can help better explain diffusion than either approach alone. They also suggest alternative options as either or both of the approaches can be used appropriate to the real situations of different application domains. We expect that our proposed approaches provide ways to quantify and understand cross-population diffusion trends at a macro level. Also, they can be applied to a wide range of research areas such as social science, marketing, and even neuroscience, for estimating dynamic influences among target regions or systems.