Discovery and analysis of social networks based on online user activities

• Main Author: Lauw, Hady Wirawan
• Other Authors: Lim Ee Peng
• Format: Theses and Dissertations
• Language: English
• Published: 2008
• Subjects: DRNTU::Engineering::Computer science and engineering::Computer applications::Social and behavioral sciences
• Online Access: https://hdl.handle.net/10356/14249
• Institution: Nanyang Technological University

This dissertation documents our research on the discovery and analysis of social networks based on user activities online, focussing on two types of data that have not previously attracted much work, namely: spatio-temporal data and collaborative rating data. Spatio-temporal data has spatial and temporal information about users. Our research aim for spatio-temporal data is to discover the latent social network that relates users together. We adopt spatio-temporal co-occurrence as a basis to predict social associations, reasoning that frequent co-occurrences by several users in space and time suggest some form of interaction or association. Our proposed model STEvent works by mining events, which are distinct co-occurrences among two or more actors, and inferring links between pairs of actors who participate in many common events. The effectiveness of the STEvent model is verified through experiments on two real-life datasets. Collaborative rating data involves a set of users (acting as reviewers) who collaboratively evaluate a set of objects, by contributing rating scores to those objects. Rating data has a bipartite social network representation, with reviewers and objects as nodes, and ratings as links. Our research agenda on rating data is to analyze a set of rating-related behaviors of reviewers and objects, focussing on three research issues involving rating-related behaviors, namely: (a) analyzing score deviations to determine the bias of reviewers and the controversy of objects, (b) analyzing score inflations/deflations to determine the quality of objects and the leniency of reviewers, and (c) analyzing score correlations to determine the rating dependencies of reviewers and objects. A common thread across these three is the notion of mutual dependency between reviewers' and objects' behaviors. For each, we propose a model that takes into account this mutual dependency, and develop the necessary computational framework to simultaneously solve the behaviors of all reviewers and objects in the network. These models are shown to be effective through various experiments on real-life and synthetic datasets.