Group Decision Making and Preference Learning on Social Networks

This thesis focuses on exploiting the dynamics and correlations of preferences over social networks for developing efficient group decision making systems. One of the main challenges in any group decision problem is learning the individual preferences upon which decisions are based. It is the position of this thesis that social networks---by capturing preference correlations across individuals induced by social interactions---provide a natural and informative platform for preference learning. By mathematical modelling of preference dynamics and correlations over social networks, we focus on developing efficient algorithms for group decision making and recommendations, with less required user data, and lower cognitive and communication burden. We introduce empathetic frameworks in which individuals derive utility based on both their own intrinsic preferences (or happiness) and empathetic preferences, determined by the satisfaction of their acquaintances. After theoretically analysing this framework, we develop a scalable algorithm for group recommendation, and empirically demonstrate its performance. To capture the correlation of preference rankings on social networks, we introduce a network formation model called ranking networks in which the similarity of two individuals' rankings determines the chance they are connected to each other. After a thorough theoretical analysis, we use a special instance that we call preference-oriented social networks, for group decision making when faced with missing preferences. We develop algorithms to predict unknown individual preferences given known preferences of others in the social network and to make effective group decisions with partial preferences. Our empirical results demonstrate that incorporating social ties can significantly improve predictions and group decision making.Ph.D