N-ary relation prediction based on knowledge graphs with important entity detection

Abstract

In recent years, the relation prediction has caught the attention of all walks of life, but previous work mainly focused on the binary relations. With the continuous expansion of real-world data, knowledge graphs have gradually expanded from binary relations to more complex higher-order relations. In particular, in n-ary facts, the influence of entities on different relations is diverse, and not all entities are indispensable. For this motivation, we propose a novel hypergraph and attention-based important entities graph convolutional network model for n-ary relation prediction. The hypergraph network considers the frequency of entities that appear in relations. Then, the relation-entity attention network assigns weights to the entities corresponding to each relation. Finally, graph convolutional network is to aggregate important higher-order information, and each layer emphasizes the influence of important entities from different perspectives. To verify the effectiveness of our algorithm, we conduct ablation analysis as well as parameters analysis on each module, and compare our algorithm with the state-of-the-art baselines. Experimental results on various datasets demonstrate the advantages of our algorithm.