Towards Multimedia-Based Storytelling in Online Social Networks

Abstract

Human activities can now be captured in real-time using sensor technology. The growth in sensor applications and smart mobile phones that come equipped with built-in sensors has led to the integration of sensors with social networks. These days, people are heavily dependent on online social networks (OSNs); they migrate their real-life activities online through various types of multimedia such as photos, videos, text, etc., which turns OSNs into a soft-sensory resource about users' events. The users use these forms of multimedia to tell their friends about their daily lives. This social network data can be crawled to build personal context-aware stories about individuals. However, the number of social users and the quantity of multimedia that is produced on social media are both growing exponentially, which leads to the challenge of information overload on OSNs. The information needed for stories, such as events and their locations, is not fully available on user's own profile. It is true that part of the information can be retrieved from the user's timeline, but a large number of events and related multimedia information is only available on friends' profiles. In this thesis, we focus on identifying a subset of close friends in order to enrich the content of the story. The amount of time people spend together has been proven to play a key role in determining close ties between people. We propose a DST (Days Spent Together) algorithm to find a user's closest friends based on the days they spent together interacting face-to-face. With the closest friends information, we are able to find additional information to complement what was found on the user's own profile, as well as to personalize the stories to ensure that they are only about the users and their closest friends. Due to the possibility of multimedia (photos in this thesis) overload for events, we propose to use the duration of events measured by DST, to determine the number of representative photos for each event. Our experiments show that the proposed approach could recognize the close friends of users and rank them from the strongest to the weakest. The results also show that with the proposed method we get days-spent-together values that are close to the corresponding true values provided by users.