RankT: Ranking-Triplets-based adversarial learning for knowledge graph link prediction

Aiming at completing the missing edges between entities in the knowledge graph, many state-of-the-art models are proposed to predict the links. Those models mainly focus on predicting the link score between source and target entities with certain relations, but ignore the similarities or differences of the whole meanings of triplets in different subgraphs. However, the triplets interact with each other in different ways and the link prediction model may lack interaction. In other word, the link prediction is superimposed with potential triplet uncertainties. To address this issue, we propose a Ranking-Triplet-based uncertainty adversarial learning (RankT) framework to improve the embedding representation of triplet for link prediction. Firstly, the proposed model calculates the node and edge embeddings by the node-level and edge-level neighborhood aggregation respectively, and then fuses the embeddings by a self-attention transformer to gain the interactive embedding of the triplet. Secondly, to reduce the uncertainty of the probability distribution of predicted links, a ranking-triplet-based adversarial loss function based on the confrontation of highest certainty and highest uncertainty links is designed. Lastly, to strengthen the stability of the adversarial learning, a ranking-triplet-based consistency loss is designed to make the probability of the highest positive links converge in the same direction. The ablation studies show the effectiveness of each part of the proposed model. The comparison of experimental results shows that our model significantly outperforms the state-of-the-art models. In conclusion, the proposed model improves the link prediction performance while discovering the similar or different meanings of triplets.