Tensor Decomposition for Link Prediction in Temporal Knowledge Graphs

We study temporal knowledge graph completion by using tensor decomposition. In particular, we use Candecomp/Parafac decomposition to factorize a given four dimensional sparse representation of a temporal knowledge graph into rank-one tensors that correspond to entities (subject and object), relations and timestamps. Using the factorized tensors, we can perform link and timestamp prediction. We compared our approach against the state of the art and found out that we are highly competitive. We report our preliminary experimental results on 5 different datasets.