ran do

M7: Large Margin Taxonomy Embedding with an Application to Document Categorization
Kilian Q. Weinberger Olivier Chapelle

m!

u Yo ll wi

an

tte

Embedding of text documents and a class taxonomy into a single joint semantic Euclidean space.
Class Taxonomy Joint low dimensional semantic space

lov

r th

High dimensional document space

ei t!

Be

classes

embedded class prototypes and documents

original documents

We derive a single convex optimization problem that learns both the regressor and class prototype positions such that Euclidean Distances satisfy large margin guarantees. We evaluate with a nearest neighbor classifier.