Examination Committee

Prof Dik Lun LEE, CSE/HKUST (Chairperson)
Prof James SHE, ECE/HKUST (Thesis Supervisor)
Prof Wenwu ZHU, Department of Computer Science and Technology, Tsinghua University (External Examiner)
Prof Peng CUI, Department of Computer Science and Technology, Tsinghua University (External Examiner)
Prof Wai Ho MOW, ECE/HKUST
Prof Danny TSANG, ECE/HKUST
Prof Bing-yi JING, MATH/HKUST

Abstract

In the era of social media, we likely interact with others online and share digital images. Many social computing applications, such as product recommendation, news virality prediction and digital marketing, make use of social graphs to identify similar users (e.g., users has similar interests or affiliations) for big data processing and information inference. However, the social graphs of users in many social media platforms may not be accessible due to various business and privacy factors. On the contrary, the widely accessible user-shared images could be an effective alternative to discover user connections for comparable social computing applications.

With over 10 million user-shared images collected from 11 social media platforms in our research, a phenomenon is proved that users with online social relationships tend to share more visually similar images. This thesis introduces a novel way to detect social signals from low level visual features and produce machine-generated labels by various multimedia computing techniques to discover user connections. This phenomenon is also proven to exist in many types of the social networks, and some novel analytic frameworks are proposed to measure, formulate and utilise this phenomenon for various common social and multimedia computing applications. To the best of our knowledge, this research work is the first attempt to use deep learning for social signal detections in social and multimedia big data for discovering user connections. Meanwhile, some real-world technology transfers and industry impacts are also realised from the results of this research work.