Machine learning on knowledge graphs

A slightly less philosophical question is How do we take this relational knowledge that we have—in knowledge graphs and relational databases, and scattered about on the internet—and use it in machine learning contexts?

One popular model for this task is the relational graph convolutional network, or R-GCN. I played a small part in the original paper, and we recently wrote a follow-up that reproduces the original, and tries to epxlain the basic principles more clearly.

A particularly promising line of research combines relational data with multimodal data. In a multimodal knowledge graph we can express subsymbolic and relational knowledge in one well-defined framework, potentially allowing ML models to learn hands-free on everythign we know in a domain. We wrote a position paper on this idea. We have several researchers currently working on this direction. Most recently we’ve published a benchmark set of knowledge graphs for evaluating this exact task (as well as relational node classification in general).