Relation prediction in knowledge graph by Multi-Label Deep Neural Network

Knowledge graph will be usefull for the intelligent system. As the relationship prediction on the knowledge graph becomes accurate, construction of a knowledge graph and detection of erroneous information included in a knowledge graph can be performed more conveniently. The goal of our research is to predict a relation (predicate) of two given Knowledge Graph (KG) entities (subject and object). Link prediction between entities is important for developing large-scale ontologies and for KG completion. TransE and TransR have been proposed as the methods for such a prediction. However, TransE and TransR embed both entities and relations in the same (or different) semantic space(s). In this research we propose a simple architecture model with emphasis on relation prediction by using a Multi-Label Deep Neural Network (DNN), and developed KGML. KGML embeds entities only; given subject and object are embedded and concatenated to predict probability distribution of predicates. Since the output of KGML is the probability distribution in [0, 1], output can be classified as positive and negative by using the threshold of 0.5. Since the output of the existing method TransE is the score in [0, ∞), the threshold value must be calculated each time. Experimental results showed that predictions by KGML are more accurate than those by TransE and TransR. KGML is more accurate than DKRL which uses both KG triples and entity descriptions for learning. KGML is more accurate than PTransE in and its learning speed is faster than PTransE. The code of KGML is available at https://github.com/yo0826jp/KGML.