Modeling and Influencing Online User Behavior with Machine Learning
Open Access
- Author:
- Zhang, Jiasheng
- Graduate Program:
- Information Sciences and Technology
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- July 05, 2020
- Committee Members:
- Dongwon Lee, Dissertation Advisor/Co-Advisor
Dongwon Lee, Committee Chair/Co-Chair
Zhenhui Li, Committee Member
Xiang Zhang, Committee Member
David Jonathan Miller, Outside Member
Mary Beth Rosson, Program Head/Chair - Keywords:
- Online User Behavior
Modeling and Influencing
Machine Learning - Abstract:
- The Internet reshapes the way people connect to the world. The online platforms such as social media and e-commerce websites collect a large number of various users' behavior data (e.g., retweets in Twitter or orders in Amazon). Such data provides an unprecedented opportunity to apply machine learning methods to modeling user behavior and further influencing it to improve the way people connect to the world. In this thesis, several specific real-world scenarios related to modeling and influencing user behavior are studied using various machine learning methods. On the other hand, challenges and observations from those scenarios inspire improvement on general machine learning methods. First, a challenge from representation of user behavior is studied. When emotional reactions of users toward posts in social media (i.e., Facebook) are represented as a ranking over a given set of emoticons, the task to predict the users' reactions given post content can be formulated as a label ranking problem. On the other hand, the imbalance property in emotional reaction data requires robustness in both label ranking performance measure and algorithms. Second, what influences user behavior is studied. More specifically, the task is to find the influence of news channels on their readers' reactions besides the content they post. It can be formulated as a multi-task learning problem. On the other hand, an observation, that the influence from channels can be different for different news content, inspires a novel multi-task learning architecture. Third, how to influence user behavior is studied. More specifically, an industry problem, recommending search story to improve search experience of users, is solved. In order to address the cross-channel challenge, a reinforcement learning framework is designed. Finally, this work is closed by a discussion on a future direction, that what users can do to combat the influence of the systems powered by machine learning methods. Recommendation problem is used as a playground to illustrate its necessity.