Feature enhancement based on hierarchical reconstruction framework for inductive prediction on sparse graphs

Knowledge graph completion aims to infer the missing links of new elements, however, the missing links often lie in sparse regions of the graph. Primary subgraph-based methods rely heavily on structural information, which makes it difficult for them to play an essential role in sparse graph completion. To address this challenge, we propose a learning framework for feature-enhanced hierarchical reconstruction (FEHR). The proposed FEHR explores relational semantics at the global and local levels, minimizing the limitations of sparse structures. First, entity graphs are converted into relation graphs, and overreliance on the entity structure is reduced by obtaining prior knowledge on similar global graphs. Second, the relational features are further refined at the local level. Finally, an improved performer model expresses the degree of preference between the predicted behaviors and relations. Extensive inductive experiments showed that FEHR performs better than state-of-the-art baselines, achieving improvements in area under the prediction–recall curve (AUC-PR) and Hits@n metrics, ranging from 0.32% to 11.73%.