Reproduce “Learning Hierarchy-Aware Knowledge Graph Embeddings for Link Prediction” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhanqiu Zhang, Jianyu Cai, Yongdong Zhang, Jie Wang

Hierarchy-Aware Knowledge Graph Embedding (HAKE) separates hierarchical position from within-level identity by representing entities with radial and angular components. Without external type annotations or ontology preprocessing, it improves link prediction over strong baselines, reaching MRRs of 0.497 on WN18RR, 0.346 on FB15k-237, and 0.545 on YAGO3-10.

Key insights

  • Radial scaling alone captures substantial hierarchical structure, but it cannot adequately distinguish entities occupying similar levels; angular information is needed for within-level separation.
  • Relations with modulus values near one tend to connect entities at similar levels, whereas systematic scaling above or below one reflects movement across hierarchy levels.
  • The clearest gains occur on datasets with stronger type-like or hierarchical structure, while improvements are smaller when relations are less semantically hierarchical.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.