10 février 2011 de 11h à 12h : salle 25-26/101
We consider the problem of labeling nodes in a multi-graph where the nodes may be connected with different types of relations. This type of problem occurs in many situations like for example the analysis of social networks or bibliographic data. Relations may be provided (e.g. friends) or computed (e.g. similarity). We propose a new learning algorithm for exploiting the rich multi-relational information in the labeling task. This method is able to optimally learn combining the influence of different relation types. It is one of the very first algorithms able to handle multi-graph information for this classi fication task. We describe experiments on four datasets which show the model ability to deal with complex relationships and to take bene fit of multi-relational information for complex labeling problems.