Analogical Inference Enhanced Knowledge Graph Embedding

Knowledge graph embedding (KGE), which maps entities and relations in a knowledge into continuous vector spaces, has achieved great success predicting missing links graphs. However, graphs often contain incomplete triples that are difficult to inductively infer by KGEs. To address this challenge, we resort analogical inference propose novel general self-supervised framework AnKGE enhance KGE models with capability. We an object retriever retrieves appropriate objects from entity-level, relation-level, triple-level. And AnKGE, train analogy function for each level of the original element well-trained model as input, outputs embedding. In order combine inductive capability enhanced interpolate score base introduce adaptive weights prediction. Through extensive experiments on FB15k-237 WN18RR datasets, show achieves competitive results link prediction task well performs inference.