A Dynamic Knowledge Graph Embedding Framework Based on Adaptive Relation Spaces and Matrix Factorization Techniques

Abstract

Knowledge graph is the knowledge of the world which provides strong support to the applications of artificial intelligence. The knowledge graph is composed of a head entity, tail entity and relation that is called as triplets. The entities and relationships of the knowledge graph give the information about the neighborhood. Over the past few years, knowledge graph embedding has proven to be indispensable in enhancing intelligent applications based on data. In this paper, a dynamic knowledge graph embedding framework with adaptive relation spaces and matrix factorization is proposed to enhance link prediction and triplet classification. A knowledge graph has triples (head, relation, tail) and entities and relations are represented in the vectors spaces. The suggested approach proposes dynamic mapping matrices created based on projection vectors, which makes the computational complexity lower because of the absence of multiplication of matrices. Benchmark datasets FB15K-237 (237 relations, 14,541 entities) and WN18RR (11 relations, 40,943 entities) are used to conduct experiments. Findings indicate a better performance, Mean Rank is lower to 199 (filtered) on FB15K-237 than on TransE, 243, and Hits10 is higher to 47.1. In WN18RR, Hits@10 is 71.9 and it is better than baseline models. In the triplet classification, the accuracy is better at 86.4% (FB15K-237) and 88.7% (WN18RR). N-to-N relation prediction is also improved by the model by about 10.1%. Also, the training time is minimized in comparison to the current models and only 24 seconds are needed instead of 70 seconds in TransR. In general, the framework is more accurate with fewer parameters and lower-cost to compute.