LatentMulti-groupMembershipGraphModel-Abstract.PDF

Latent Multi-group Membership Graph Model Myunghwan Kim MYKIM@STANFORD.EDU Stanford University, Stanford, CA 94305, USA Jure Leskovec JURE@CS.STANFORD.EDU Stanford University, Stanford, CA 94305, USA Abstract that it represents. Above models can ?nd patterns which We develop the Latent Multi-group Membership account for the connections between nodes, but they can- Graph (LMMG) model, a model of networks not account for the node features. with rich node feature structure. In the LMMG Node features along with the links between them provide model, each node belongs to multiple groups and rich and complementary sources of information and should each latent group models the occurrence of links be used simultaneously for uncovering, understanding and as well as the node feature structure. The LMMG exploiting the latent structure in the data. In this respect, we can be used to summarize the network structure, develop a network model that considers both the emergence to predict links between the nodes, and to pre- of links of the network and the structure of node features dict missing features of a node. We derive ef?- such as user pro?le information or text of a document. cient inference and learning algorithms and eval- uate the predictive performance of the LMMG on Considering both sources of data, links and node features, several social and document network datasets. leads to more powerful models than those that only con- sider links. For example, given a new node with a few of its links, traditional network models provide a predic- 1. Introduction tive distribut