Inductive link prediction aims to predict missing triplets involving newly emerged entities in the Knowledge Graphs (KGs) of the inference phase. Despite advancements in this field, inferring comparative relationships for unseen entities with numeric ...
Inductive link prediction aims to predict missing triplets involving newly emerged entities in the Knowledge Graphs (KGs) of the inference phase. Despite advancements in this field, inferring comparative relationships for unseen entities with numeric attributes remains challenging, primarily due to variations in numeric distributions. Moreover, the inherent asymmetric nature of numeric comparison relations (e.g., ”is_taller_than”) complicates the direct application of existing inductive link prediction and numeric encoding models, which often fail to distinguish between such asymmetric pairs. To address these challenges, we propose Relation-aware Relative Numeric Encoding (RRNE), a novel approach enhancing inductive numerical reasoning by considering the relative nature of numerical reasoning. Our method effectively mitigates the aforementioned challenges by computing relative differences between each entity in the subgraph and both the head and tail of the target triplet at the representation level. As a result, the entity representations become more robust when the inference KG exhibits a numerical distribution that is unseen yet follows a similar pattern during training. Also, our method leads to effectively capturing the asymmetry inherent in numeric comparison relations. We conduct extensive experiments on three benchmark datasets―Credit, Spotify, and US-Cities―for inductive numerical reasoning. Our empirical results demonstrate that RRNE significantly outperforms existing baselines, achieving enhancements of up to 24.81% in AUC-PR and 20.69% in AUC. These outcomes validate the effectiveness of our approach in tackling inductive numerical reasoning.