User Attribute Learning from Social Media and Ubiquitous Sensors

Mr Liu Ye
Dr Rosenblum, David S., Provost'S Chair Professor, School of Computing

  24 Apr 2018 Tuesday, 10:00 AM to 11:30 AM

 Executive Classroom, COM2-04-02


With the growth of the ubiquitous sensors and social media services, there are numerous services available for users across different electronic devices and social services, such as wearable sensors, smartphone, tablet computers, various social networks. An increasing number of individuals are immersed in those services and are generating a huge amount of user-generated data in their daily lives. Those huge amount of digital data generated from users online behaviors contains rich information about users various knowledge, such as gender, age, religion, geographic location, interests, education background and political preference. The availability of such user attributes are crucial for many useful applications like personalization, advertising and recommendation. In this context, how can we learn users' various attributes from those data has become an indispensable task, which can help understand users' interests more accurately, and provide more personalized services and greatly facilitate the online targeting as well as advertising.

Nowadays, users are involved in both physical world and digital environment simultaneously. Hence, there are attributes that can depict either the digital identity or the physical identify of a given user. Those attributes across the physical and digital worlds can be denoted as physical and digital attributes, where the digital attributes can be further categorized as: static digital attributes and dynamic digital attributes. In this thesis, we will investigate three representatives of these broader classes, and develop the approaches that can be easily generalized to the broader class of user attribute learning.

This thesis first investigates the user physical attribute learning and focus on the case study of user activity attribute learning from multiple sensors. As various sensors are acting as the very important component of users' lives in physical world, learning user attributes from multiple sensors is generalizable to their physical attribute learning. The thesis then studies the user dynamic digital attribute learning and targets at the case study of user occupational attribute learning from multiple social networks. In particular, it presents a novel multi-task multi-source learning approach to capture the dynamic progression in user occupational stages, which can be easily generalized to other dynamic digital attribute learning scenarios. Finally, the thesis also presents a novel multi-task learning framework to learn user political attributes as a case study of static digital attribute learning, where the developed approaches can be adapted to other static attribute learning applications.

To evaluate the effectiveness of the presented approaches, several real-world datasets have been built and extensive experiments have been conducted on the these datasets. The experimental results have demonstrated that our approaches could obtain significant gains in learning user attributes from ubiquitous sensors and social media.