Open-world knowledge graph completion for unseen entities and relations via attentive feature aggregation

Most of knowledge graph completion (KGC) models are designed for static KGs where entity and relation sets are fixed. These approaches are inherently transductive because they simply predict the plausibility of facts whose entities and relations had to previously appear in the training phase. However, since entities and relations are constantly added, removed, or changed over time, KGC models should be able to generalize to out-of-KG entities and relations in evolving KGs, which are more suited to real-world scenarios. Moreover, incorporating global graph-structured information into KGC models is another challenging issue. To overcome these issues, this paper proposes a novel Inductive KG Embedding (IKGE) model for open-world KGC, which accommodates out-of-KG entities and relations. Unlike training individual unique embeddings, the proposed model fundamentally learns an embedding generator function to elaborately generate fact embeddings in an inductive manner. Specifically, the feature information of each fact is extracted as a vector from its entity-related and relation-related side information via an attention mechanism. Then, to score a given fact, our neighborhood feature aggregator hierarchically accumulates the feature information of multi-hop neighbors. Experimental results show that IKGE outperforms existing approaches in both transductive and inductive setups by successfully aggregating neighborhood features.