A Relation Prediction Method for Industrial Knowledge Graph with Complex Relations

Abstract

In the industrial knowledge graph, relations between entities has important physical significance. In some cases, the relations determines product quality or fault prediction accuracy. Therefore, relation prediction is particularly important. The existing translation models are poor in relation prediction, because the current method of generating incorrect triples cannot generate incorrect triples with incorrect relations. Therefore, this paper develops an incorrect triple generation method by combining uniform distribution and Bernoulli distribution to improve the relation prediction accuracy. Then UBTransH model is developed by combining TransH with the incorrect triples based on uniform distribution and Bernoulli distribution. In the UBTransH, the relation replacement probability and entity replacement probability of the correct triples are first obtained through uniform distribution, then the head entity replacement probability and tail entity replacement probability are calculated with the entity replacement probability and Bernoulli distribution. Second, the incorrect triples are generated by replacing the relation/head entity/tail entity with the relation/head entity/tail entity replacement probability. Third, the incorrect triples are applied to train TransH. Finally, based on relation prediction, the developed UBTransH are compared with TransH on several typical datasets. Experimental results show that the developed UBTransH significantly improves the relation prediction accuracy.