Evaluating Incremental Embeddings for Evolving Knowledge Graphs

Incremental knowledge graph embedding is the task of inferring a representation from knowledge graph so that (1) plausible triples are scored higher than implausible ones and for which additionally (2) the representation can be updated to incorporate changes in the knowledge,e.g., the deletion of existing triples or the insertion of new ones. In this talk, we will survey existing approaches for this task and demonstrate that they have not sufficiently evaluated so far for their incremental capabilities. Thereafter, we will propose a new comprehensive evaluation scheme for incremental knowledge graph embeddings and discuss what remains to be investigated.