责任导师 : Patrick GALLINARI

助理责任导师 : PIWOWARSKI Benjamin

Inference of signed links in social networks, by learning from user interactions

This thesis deals with the analysis of relationship polarity in social networks, using statistical learning methods. Many networks on Internet explicitly or implicitly contain bivalent relationships (such as friend / enemy). Such relationships can be exploited in many tasks, such as recommendation or content suggestion. The study of the polarity of social relations raises many fundamental problems: the positive and negative relationships between individuals have different semantics and few formal machine learning models exploit signed links.
In this thesis, we are specifically interested in implicitly signed links, inferred from users judgments on items. We propose several models that exploit the tasks of recommendation and link prediction. Our main hypothesis is that agreements between users do not have the same semantics depending on whether they are linked by a positive or negative link. We first propose a collaborative filtering model for recommendation that exploits the polarity of agreements between neighbors. We then develop user models to refine the characterization of the relationships between users by taking into account their propensity to rate positively or negatively. Finally, we explore the use of agreements polarity to predict the sign of social connections between users.

2014-2015 刊物