(A) topic-based inter-region network for spatio-temporal analysis of large SNS data

Abstract

Understanding interest divergence over the regions allows a novel analysis of SNS data such as global consensus and grouping regions. In dissertation, we propose the notion of topic versatility that refers to the degree to a topic's coverage for different points of view or sub-topics over the regions in social media. Since versatility of a topic captures its characteristics across different regions, identifying versatile topics can help understand regional interests. In addition, it becomes possible to identify regions with minority sub-topics, which can be easily obscured. Then, we propose the set of methods to analyze the dynamics of interests among different regions for the specific topic. In particular, we generate a topic-based inter-region networks and cluster regions based on sub-topic interests. We apply Markov Chain Clustering (MCL) to group regions for their commonality in sub-topic interest. The clustering results are further used to reveal the level of consensus and dissensus among the regions. For illustration, we show analyses of four topics, two for the cases where versatility value drops and the other for increases. Finally, we use the inter-region network to conduct information propagation analysis and event mention prediction. We propose the edge weighting schemes for accurate propagation modeling, and show that the graph with the varying weighted sum of follower network and topical similarity-based network for edge weighting gives the best performances. For the second task, we propose a method for predicting major event mentions in Tweets based on the multivariate regression method utilizing the regional event occurrence status. In particular, we apply region clusters that noisy relationships among the regions can be removed. Compared to the state-of-art method for predicting events, we show the proposed method of using the topical similarity-based network of regions gives a better performance.