Notes on: Trouillon, Th\'eo, Welbl, J., Riedel, S., Gaussier, \'Eric, & Bouchard, G. (2016): Complex embeddings for simple link prediction

Table of Contents

Overview

  • Relational models ought to be able to
    1. Learn:
      • Reflexivity / irreflexivity
      • Symmetry / antisymmetric
      • Transistivity
    2. be linear in both time and memory to scale
  • Standard embedding-approaches satisfy the above but has a problem: anti-symmetry blows up the parameter space
  • Attempts to solve the concern above using complex embeddings, which can model anti-symmetry of relations