Detecting hierarchical relationships and roles from online interaction networks

In social networks, analysing the explicit interactions among users can help in inferring hierarchical relationships and roles that may be implicit. In this thesis, we focus on two objectives: detecting hierarchical relationships between users and inferring the hierarchical roles of users interacting via the same online communication medium. In both cases, we show that considering the temporal dimension of interaction substantially improves the detection of relationships and roles. The first focus of this thesis is on the problem of inferring implicit relationships from interactions between users. Based on promising results obtained by standard link-analysis methods such as PageRank and Rooted-PageRank (RPR), we introduce three novel time-based approaches, “Time-F” based on a defined time function, Filter and Refine (FiRe) which is a hybrid approach based on RPR and Time-F, and Time-sensitive Rooted-PageRank (T-RPR) which applies RPR in a way that takes into account the time-dimension of interactions in the process of detecting hierarchical ties. We experiment on two datasets, the Enron email dataset to infer managersubordinate relationships from email exchanges, and a scientific publication coauthorship dataset to detect PhD advisor-advisee relationships from paper co-authorships. Our experiments demonstrate that time-based methods perform better in terms of recall. In particular T-RPR turns out to be superior over most recent competitor methods as well as all other approaches we propose. The second focus of this thesis is examining the online communication behaviour of users working on the same activity in order to identify the different hierarchical roles played by the users. We propose two approaches. In the first approach, supervised learning is used to train different classification algorithms. In the second approach, we address the problem as a sequence classification problem. A novel sequence classification framework is defined that generates time-dependent features