Abstract
Most existing knowledge graphs suffer from incompleteness, which can bealleviated by inferring missing links based on known facts. One popular way toaccomplish this is to generate low-dimensional embeddings of entities andrelations, and use these to make inferences. ConvE, a recently proposedapproach, applies convolutional filters on 2D reshapings of entity and relationembeddings in order to capture rich interactions between their components.However, the number of interactions that ConvE can capture is limited. In thispaper, we analyze how increasing the number of these interactions affects linkprediction performance, and utilize our observations to propose InteractE.InteractE is based on three key ideas -- feature permutation, a novel featurereshaping, and circular convolution. Through extensive experiments, we findthat InteractE outperforms state-of-the-art convolutional link predictionbaselines on FB15k-237. Further, InteractE achieves an MRR score that is 9%,7.5%, and 23% better than ConvE on the FB15k-237, WN18RR and YAGO3-10 datasetsrespectively. The results validate our central hypothesis -- that increasingfeature interaction is beneficial to link prediction performance. We make thesource code of InteractE available to encourage reproducible research.