Reviews: Expanding Holographic Embeddings for Knowledge Completion

They define a model which generalises an existing low-complexity model HolE by stacking a number of instances of HolE, each perturbed with a perturbation vector c. The authors show how, for an appropriately chosen set of c vectors, this model is equivalent to RESCAL, a high-complexity model. They provide a number of theoretical results characterising their model for two different classes of perturbation vectors. Finally, they demonstrate that their model improves on existing methods on the FB15K dataset.