Collective Social Activity Prediction by Using Continuous-time Stochastic Process

Open Access
- Author:
- Huang, Shu
- Graduate Program:
- Information Sciences and Technology
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- April 25, 2012
- Committee Members:
- Dongwon Lee, Thesis Advisor/Co-Advisor
- Keywords:
- social network
social activity
prediction
algorithm - Abstract:
- In recent years, online social network sites have successfully emerged to attract a huge number of users. Due to the tremendous number of users, these sites are playing a more and more important role in business marketing and customer care. The ex-ante knowledge of the future social activity and network status can assist a lot in decision making and improve the profit of businesses. Existing approaches analyze social networks and predict their future behaviors by modeling the evolving network with all cumulative nodes (members) and connections. Although the arrival of new nodes is taken into consideration to address the dynamic nature of social networks, in general, the degree of fine-grained node activity (e.g. a user adds a friend or posts on a friend's wall) has played very little role in such an analysis. As the infrastructure of a social network heavily depends on the member activities, we propose to explore the collective social dynamics and community infrastructure to predict the future social activity. As the evolution of social networks is a random process, we make use of a continuous-time stochastic process model with an uncountable future state space to simulate the social dynamics and investigate the impact of user activities. In this thesis, we derive a novel parameterized model that incorporates the information embedded in collective activities to predict various features of the social community, including the size of active population, the number of social interactions, the social sentiment trend, etc. With member activities evolving over time, the predicting model itself also evolves and therefore dynamically simulates the network status to fit the real-time characteristics of the current active population. Our experiments using two real social media datasets (Facebook and CiteSeer) show that the proposed parameterized stochastic model is effective to simulate the social activity evolution and can predict the collective social activities with more than 80% accuracy through different time scales.