Relation-Aware Multi-View Representation Learning for Cross-KG Relation Prediction
Abstract
Cross-KG relation prediction aims to predict possible relations between entities from different knowledge graphs. This task is difficult because different graphs often have different structures, use different relation types, and share only a small number of entities. Under this setting, it is not easy to pass useful information across graphs. In this paper, we propose a relation-aware multi-view representation learning framework for cross-KG relation prediction. We first build a unified graph by connecting two knowledge graphs through overlapping entities. Based on this graph, we learn entity representations from two complementary views: a structural view that captures graph topology, and a relation-aware view that uses relation information to guide message passing through a gating mechanism. The two views are then fused to obtain final entity embeddings, which are used to rank candidate cross-KG relations. Experiments on DFB-1, WK3L-E, and MEDICINE show that our method consistently outperforms competitive baselines, including the dedicated cross-KG model CrossKG. In particular, compared with CrossKG, our method reduces MR by 65.1%, 5.7%, and 33.7% on the three datasets, respectively. Ablation and overlap-ratio experiments further show that the two views work well together, and that relation-aware propagation is especially helpful when the overlap between graphs is low.