The rise in the importance of social media platforms as communication tools has been both a blessing and a curse. For scientists, they offer an unparalleled opportunity to study human social networks. However, these platforms have also been used to propagate misinformation and hate speech with alarming velocity and frequency. The overarching aim of our research is to leverage the data from social media platforms to create and evaluate a high-fidelity, at-scale computational simulation of online social behavior which can provide a deep quantitative understanding of adversaries' use of the global information environment. Our hope is that this type of simulation can be used to predict and understand the spread of misinformation, false narratives, fraudulent financial pump and dump schemes, and cybersecurity threats. To do this, our research team has created an agent-based model that can handle a variety of prediction tasks. This dissertation introduces a set of sampling and deep learning techniques that we developed to predict specific aspects of the evolution of online social networks that have proven to be challenging to accurately predict with the agent-based model. First, we compare different strategies for predicting network evolution with sampled historical data based on community features. We demonstrate that our community-based model outperforms the global one at predicting population, user, and content activity, along with network topology over different datasets. Second, we introduce a deep learning model for burst prediction. Bursts may serve as a signal of topics that are of growing real-world interest. Since bursts can be caused by exogenous phenomena and are indicative of burgeoning popularity, leveraging cross-platform social media data is valuable for predicting bursts within a single social media platform. An LSTM model is proposed in order to capture the temporal dependencies and associations based upon activity information. These volume predictions can also serve as a valuable input for our agent-based model. Finally, we conduct an exploration of Graph Convolutional Networks to investigate the value of weak-ties in classifying academic literature with the use of graph convolutional neural networks. Our experiments look at the results of treating weak-ties as if they were strong-ties to determine if that assumption improves performance. We also examine how node removal affects prediction accuracy by selecting nodes according to different centrality measures. These experiments provide insight for which nodes are most important for the performance of targeted graph convolutional networks. Graph Convolutional Networks are important in the social network context as the sociological and anthropological concept of 'homophily' allows for the method to use network associations in assisting the attribute predictions in a social network.
If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu.
Doctor of Philosophy (Ph.D.)
College of Engineering and Computer Science
Length of Campus-only Access
Doctoral Dissertation (Open Access)
Hajiakhoond Bidoki, Neda, "Stochastic Sampling and Machine Learning Techniques for Social Media State Production" (2020). Electronic Theses and Dissertations, 2020-. 225.