Human beings are born social. In the social media era, we share and interact with others digitally, forming online social graphs and sharing billions of images. Many social media applications, such as recommendation, virality prediction, and marketing, make use of social graphs as similar users (e.g., users with similar interests) tend to be friends. However, the social graph may not be explicitly specified by users or may be kept private due to privacy concerns. Meanwhile, billions of user-shared images are shared by individuals, and the images are widely accessible to others due to their sharing nature. These user shared images are proved to be a more effective alternative to discover user connections. This thesis introduces a novel way to detect social signals from low leve...[ Read more ]

Human beings are born social. In the social media era, we share and interact with others digitally, forming online social graphs and sharing billions of images. Many social media applications, such as recommendation, virality prediction, and marketing, make use of social graphs as similar users (e.g., users with similar interests) tend to be friends. However, the social graph may not be explicitly specified by users or may be kept private due to privacy concerns. Meanwhile, billions of user-shared images are shared by individuals, and the images are widely accessible to others due to their sharing nature. These user shared images are proved to be a more effective alternative to discover user connections. This thesis introduces a novel way to detect social signals from low level visual features, and to represent them with unbiased machine-generated labels to discover user connections. Based on 11 million user-shared images from 11 real social media platforms, a phenomenon exists that related users who have online friendships or follower/followee relationships on those platforms share more similar images. This phenomenon is independent of the network origins, the content sharing mechanisms and the image processing/computer vision techniques that encode the images. Hence, an analytic framework is proposed to measure, formulate and utilize the phenomenon for follower/followee recommendation. The framework is optimized for social signal detections using deep learning. Different applications are also discussed. To the best of our knowledge, this framework is the first attempt to discover connections by detecting social signal from user-shared images.