CONF dhiraj:hcm:2006/IDIAP Finding groups of people in Google news Joshi, Dhiraj Gatica-Perez, Daniel EXTERNAL http://publications.idiap.ch/attachments/papers/2006/dhiraj-hcm-2006.pdf PUBLIC http://publications.idiap.ch/index.php/publications/showcite/dhiraj:rr05-68 Related documents ACM Int. Conf. on Human-Centered Multimedia (HCM) 2006 IDIAP-RR 05-68 In this paper, we study the problem of content-based social network discovery among people who frequently appear in world news. Google news is used as the source of data. We describe a probabilistic framework for associating people with groups. A low-dimensional topic-based representation is first obtained for news stories via probabilistic latent semantic analysis (PLSA). This is followed by construction of semantic groups by clustering such representations. Unlike many existing social network analysis approaches, which discover groups based only on binary relations (e.g. co-occurrence of people in a news article,',','), our model clusters people using their topic distribution, which introduces contextual information in the group formation process (e.g. some people belong to several groups depending on the specific subject). The model has been used to study evolution of people with respect to topics over time. We also illustrate the advantages of our approach over a simple co-occurrence-based social network extraction method. REPORT dhiraj:rr05-68/IDIAP Finding groups of people in Google news Joshi, Dhiraj Gatica-Perez, Daniel EXTERNAL http://publications.idiap.ch/attachments/reports/2005/dhiraj-idiap-rr-05-68.pdf PUBLIC Idiap-RR-68-2005 2005 IDIAP Published in Proc. ACM Int. Conf. on Multimedia, Workshop on Human-Centered Multimedia (HCM,',','), Santa Barbara, Oct. 2006. In this paper, we study the problem of content-based social network discovery among people who frequently appear in world news. Google news is used as the source of data. We describe a probabilistic framework for associating people with groups. A low-dimensional topic-based representation is first obtained for news stories via probabilistic latent semantic analysis (PLSA). This is followed by construction of semantic groups by clustering such representations. Unlike many existing social network analysis approaches, which discover groups based only on binary relations (e.g. co-occurrence of people in a news article,',','), our model clusters people using their topic distribution, which introduces contextual information in the group formation process (e.g. some people belong to several groups depending on the specific subject). The model has been used to study evolution of people with respect to topics over time. We also illustrate the advantages of our approach over a simple co-occurrence-based social network extraction method.