Relational Learning Models for Social Network Analysis

Social networks have been studied for nearly half a century by sociologists to analyze interactions between people. Nowadays, with the advent of Web 2.0, social networks have moved from being an abstract concept to actual online applications such as Facebook, Twitter and Linkedin, which are used daily by people to create and maintain relationships with friends, co-workers and other acquaintances. However, online social networks allow their users to do more than just maintain friendships: people can generally create and share content in various forms, from simple textual messages (called posts) to photos, videos, audios and much more. Several approaches in Social Network Analysis have been proposed in the years to extract knowledge from social networks, addressing tasks that ranges from understanding how users create and modify their relationships, to finding the most influential people in a group, to understanding the ideas and opinions expressed by people in their posts. Many techniques from the field of Machine Learning have been used to address these problems. While some of them exploit the relationships among users, others focus on the content generated by the users, typically by analyzing the textual content written in the posts. These approaches, however, are generally unable to exploit both, in this way ignoring a consistent part of information available in social networks. The field of Relational Learning tries to overcome this limitation, by extending traditional approaches in order to use both sources of information, and thus achieve better performances. In this thesis, I propose new Relational Learning approaches that address two tasks in Social Network Analysis. The first task is Community Discovery, which objective is to detect groups of users that share strong connections (e.g. working in the same company, attended the same school, etc.) or sharing the same interests. While this task is generally addressed by considering only the network structure, adding the user content can allow to increase the performance. The second task is Opinion Detection, which objective is to infer the opinion of users about a specific topic (politics, likeness of a brand). This task is typically addressed using user textual content, but the relationships can provide additional insights that allow to improve the inference of users’ opinions. The experimental investigations reveal that network structure and user-generated content provide complementary information, and that using both sources of data can improve the performance of algorithms in both community discovery (a structure-based task) and opinion detection (a content-based task). iii