11 Knowledge graph completion

The previous chapter is an introduction to link prediction and completion techniques. The difference between link prediction and completion is that the first is a workflow to predict future links, while the latter deals with predicting missing links. However, in practice, link prediction and completion workflows are very similar. What wasn’t explicitly mentioned is that the link prediction features used in Chapter 10 do not differentiate between various node or relationship types. For example, the number of common neighbors does not differentiate between different relationship or node types. Therefore, the link prediction features used in Chapter 10 work best with monopartite or homogeneous graphs. A monopartite or a homogenous graph consists of a single node and relationship type.

Visualization in Figure 11.1 depicts a homogeneous graph that consists of a single node type Stream and a single relationship type SHARED_AUDIENCE.

11.1 Knowledge graph embedding model

11.1.1 Triple

11.1.2 TransE

11.1.3 TransE limitations

11.2 Knowledge graph completion

11.2.1 Hetionet

11.2.2 Dataset split

11.2.3 Train a PairRE model

11.2.4 Drug application predictions

11.2.5 Explaining predictions

11.3 Summary

11.4 References

11.5 Solutions to exercises