Relational Classification for Personalized Tag Recommendation

Folksonomy data is relational by nature, and therefore methods that directly exploit these relations are prominent for the tag recommendation problem. Relational methods have been successfully applied to areas in which entities are linked in an explicit manner, like hypertext documents and scientific publications. For approaching the graph-based tag recommendation task of the ECML PKDD Discovery Challenge 2009, we propose to turn the folksonomy graph into a homogeneous post graph and use relational classification techniques for predicting tags. Our approach features adherence to multiple kinds of relations, semi-supervised learning and fast predictions.