Convex Co-embedding

We present a general framework for association learn-ing, where entities are embedded in a common la-tent space to express relatedness via geometry—an ap-proach that underlies the state of the art for link pre-diction, relation learning, multi-label tagging, relevance retrieval and ranking. Although current approaches rely on local training methods applied to non-convex formu-lations, we demonstrate how general convex formula-tions can be achieved for entity embedding, both for standard multi-linear and prototype-distance models. We investigate an efficient optimization strategy that al-lows scaling. An experimental evaluation reveals the ad-vantages of global training in different case studies.