SOCIAL NETWORK MODELING, LINK PREDICTION, AND SENTIMENT IMPACT ANALYSIS
Open Access
- Author:
- Qiu, Baojun
- Graduate Program:
- Computer Science
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- July 29, 2011
- Committee Members:
- John Yen, Dissertation Advisor/Co-Advisor
John Yen, Committee Chair/Co-Chair
Guohong Cao, Committee Chair/Co-Chair
Jesse Louis Barlow, Committee Member
Jia Li, Committee Member
Qiyang He, Committee Member
Raj Acharya, Committee Member - Keywords:
- temporal data mining
machine learning
Social network analysis
text mining
sentiment analysis
link prediction - Abstract:
- Social network dynamics analysis is one of the most important fields in social network analysis. It studies the temporal network structure and impacts on actors at both macro and micro levels. Specifically, social network modeling is the macro-study of network structure. It identifies general principles of link formation that lead to interesting network properties. Link prediction is the micro-study of network structure. It predicts future links between nodes. The impact analysis investigates impacts of links on individual nodes, and impact dynamics of the networks. Although the links between actors in social networks are the results of social interactions between the actors, most of the current network dynamics analysis is link centric. For instance, social network modeling studies networks by simulating individual links. A link centric approach for dynamics analysis has the following problems: 1) difficult to model n-ary ($n>2$) relationship; 2) difficult to include properties of social interaction in modeling network dynamics; 3) difficult to aggregate properties of past social interactions in network modeling and prediction; 4) difficult to analyze impacts of social interactions (on actors and networks) using their properties. To solve the problems, we developed new approaches for social dynamics analysis from the perspective of social interactions. Specifically, to model n-ary ($n geq 2$) social interactions and incorporate their properties in modeling networks, we investigate an event-driven social network modeling to model behavior patterns of multi-actors social interactions. To aggregate properties of past social interactions to predict future links between actors, we develop behavior evolution based link prediction to discover temporal behavior patterns for predicting future links. Finally, we use the characteristics of social interactions to investigate the sentiment impact of interactions in an online heath community. The experimental results of social network modeling suggest that our event-driven social network modeling can generate realistic networks exhibiting important macroscopic properties, such as power-law degree distribution, hierarchical community structure and assortativity which are similar to real networks. The experimental results of link prediction indicate that our behavior evolution based link prediction approach consistently achieves significant improvement on link prediction accuracy on multiple real networks. Finally, our work in sentiment impact analysis discovers the patterns of sentiment change of members of a online heath community, and identify factors that affect the sentiment change. These research results indicate the benefits of modeling social interactions directly for characterizing and predicting dynamic behaviors of social networks at both macroscopic and microscopic levels.