- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper introduces SOHL to help preserve the semi-order of entities in knowledge hypergraphs and further improve link prediction performance in hyperedge representation learning. The rationale is that permuting entities within the same logical-role group preserves the semantics of a hyperedge, while swapping entities across different groups can alter its semantics. To achieve this, the authors use a counterfactual-guided logical role mining mechanism to learn latent logical role group patterns. They evaluate their method’s performance using MRR to measure recovery from corrupted inputs. They also evaluate the algorithm’s efficiency and compare it with that of other order-preserving methods.
Critical strengths
The rationale behind this paper makes sense to me and is clearly explained. The paper is well structured and presented.
Critical weaknesses
-
Equations 13–15 do not adequately support the claimed role-aware encoding, which makes the claim unconvincing. First, there is a conflict between Equations 13 and 14. The notation [; ] is not clearly explained in the paper. Assuming that it denotes concatenation, as is common in graph learning, the dimension of should therefore be . However, Equation 14 adds , with dimension , to a weighted , with dimension , which is dimensionally inconsistent. Either the equation is incorrect or some components are missing. Second, combining Equations 13, 14, and 15 would cause the soft assignment matrix to cancel out, preventing the proposed role-learning mechanism from functioning as intended. Any additional operations needed to resolve these issues are missing from the description.
-
The experimental reporting contains unexplained discrepancies. In Table 1, HYPER achieves an MRR very close to that of SOHL, ranking second in most settings. However, in Figure 3(a), HYPER has the lowest MRR, with a large gap relative to SOHL, even when the legal permutation ratio is 0%. The paper does not explain this discrepancy. Also, the linked repository contains only a README and provides neither the implementation nor the datasets, so the implementation cannot be examined to clarify these inconsistencies.
Questions for authors
- How are the dimensions of and aligned for the addition in Equation (14)? Is a projection or another transformation omitted from the description?
- In Table 1, HYPER performs close to SOHL in most settings. Why does HYPER have the lowest MRR in Figure 3(a), even at a 0% legal permutation ratio? Do these experiments use different checkpoints or evaluation settings?