Thread by @michael_galkin

We are glad to announce that our paper "Message Passing for Hyper-Relational Knowledge Graphs" by @michael_galkin @shape_mismatch @__gauravm @Ricardo_Usbeck @JLehmann82 has been accepted at @emnlp2020! 🎆 We propose a #GNN architecture for hyper-relational KGs like @wikidata. ⬇️

Traditional KGs are based on triples, whereas new KGs like #wikidata use statements and qualifiers to instantiate each edge further making the graph hyper-relational (img1). We incorporate these qualifiers by modifying existing multi-relational GNN (CompGCN) in the StarE (img 2).

This improves link prediction performance across the board!

In fact, our architecture can be used for encoding increasingly growing RDF* graphs - kudos to @olafhartig :) This was our inspiration to name the architecture as StarE ⭐️

But wait, there's more! 🍬 We systematically analyze existing datasets and find various shortcomings including test leakages. Thus we propose a new hyper-relational dataset WD50k. We made sure the distribution of qualifiers is close to that in the whole #Wikidata.

We can do lots of stuff with that encoder now! 🚀Varying the amount of qualifiers in the KG we find (img 4):
1. The GNN encoder boosts LP drastically 💪
2. The more qualifiers - the better. In fact, including as few as 2 qualifiers can already improve the link prediction.

Shoutout to the #KnowledgeGraph community - start adding more qualifiers 😉
We plan to introduce more hyper-relational KG datasets and experiment with more #GNN tasks like node classification in transductive and inductive setups 🧙‍♂️. Stay tuned for the blog, code, and full 📜 !