acceptodds
Under review as a conference paper at ICLR 2027

KGE-Path: Interpretable Link Prediction based on Knowledge Graph Embedding.

Abstract

Explaining why a prediction is made is essential for the prediction to be trusted and acted upon. Most explainers of link predictions made by graph and knowledge-graph models are post-hoc: they treat the prediction model as a black box and perturb its input to obtain an explanation. This forces the model to run on inputs whose distribution differs from the one it was trained on, so the faithfulness of the explanation is assumed rather than guaranteed. It is also time-consuming, since an additional model has to be trained or optimized to produce the perturbation. To address both issues, we propose **KGE-Path**, a transparent knowledge graph embedding (KGE) model whose prediction relies only on the geometric relationship between the embeddings of the two nodes, together with an explanation method that is derived from this relationship. We exploit this transparency and score each path connecting the two query nodes using only the embeddings of the model and the topological structure of the input graph; the resulting paths, together with their scores, form an explanation that is read off directly from the model and therefore faithfully reflects how the prediction is made. To make the model better suited to this explanation, we also adapt the standard KGE training: an *edge penalty* that performs negative sampling over the relation types, a *node penalty* that exploits the information of the neighborhood, and a regularization term that further aligns the explanation with the prediction. On six datasets, KGE-Path explains better than GNN- and KGE-based explainers, and it explains a query pair faster than the baselines on four of the six datasets.

Primary area:
interpretability and explainable AI
Keywords:
Knowledge graph embedding, link prediction, interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_n6jMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes KGE-Path, an interpretable knowledge graph embedding framework for link prediction. It derives predictions from the geometric relationships between entity embeddings and generates path-based explanations by measuring geometric consistency. The method further introduces edge and node penalties and a regularization term to improve prediction and explanation. Experiments on six datasets demonstrate competitive prediction performance, improved explanation quality, and faster per-query explanation on four datasets.

Critical strengths
  1. Unified prediction and explanation framework. KGE-Path derives predictions and path-based explanations from the same embedding geometry, avoiding additional explainer training. This provides a coherent and potentially efficient approach to interpretable link prediction.
  2. Strong empirical performance. Experiments across six datasets demonstrate competitive prediction accuracy and consistently strong explanation results compared with existing baselines. The method also achieves faster per-query explanation on four datasets, supporting its practical value.
Critical weaknesses
  1. Insufficient justification of explanation faithfulness. Although Equation (9) establishes a geometric connection between prediction and explanation, Assumption 1 further links smaller path composition errors to stronger evidential support. This connection is not sufficiently justified or directly validated, leaving some uncertainty about the paper's claim that the explanations faithfully reflect how predictions are made.
  2. Insufficient ablation of key training components. The ablation study evaluates the overall training adaptation without separately examining the edge and node penalties. As these components serve different purposes, their individual contributions remain unclear, limiting the empirical support for the proposed training design.
Questions for authors
  1. Beyond the geometric relationship established in Equation (9), what theoretical or empirical evidence supports Assumption 1 that lower path composition errors indicate stronger evidence for the predicted relation?
  2. Could the authors provide separate ablation results for the edge penalty and node penalty to clarify their individual contributions to prediction and explanation performance?
  3. How are multiple ground-truth relations between the same ordered entity pair handled during training and evaluation, given the single-valued relation mapping in Section 2 and the prediction rule in Equation (7)?
  4. What motivates the selection of only the highest-scoring path as the explanation path in Equation (11)? Could multiple high-scoring paths provide more complete explanations, particularly when the ground truth contains multiple explanatory paths?
  5. How does KGE-Path differ from recent related methods such as PGNN-Link (IEEE Transactions on Knowledge and Data Engineering, 2026) and KGC-Explainer (IEEE Transactions on Reliability, 2026) in terms of its explanation mechanism and contributions? Discussing these recent methods in the paper would help clarify the positioning and distinctive contributions of KGE-Path.
Reviewer_7dz7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes KGE-Path, a RotatE-based model for relation prediction between a given pair of nodes. Each candidate path between the two nodes is scored by how well the composition of its relation embeddings maps the head embedding onto the tail embedding, and the best-scoring paths are returned as the explanation. The paper also changes the standard KGE training by adding an edge penalty over relation types, a neighbourhood-based node penalty, and a regulariser. It reports better explanation and prediction results than GNN- and KGE-based baselines on six datasets.

Critical strengths

Fast explanations. The proposed approach is up to ~22× faster per query pair than the fastest baseline explainer

Critical weaknesses
  1. Missing discussion of path-based link predictors; the novelty appears thin. KGE-Path scores a link by composing relation embeddings along the paths that connect the two entities, and returns the highest-scoring paths as the explanation. This is the same setting as path-based link predictors such as NBFNet [1], A*Net [2] and MINERVA [3], whose paths also provide an interpretation of the prediction; NBFNet in particular derives path interpretations by scoring and ranking the paths between the two entities. None of these works is cited. The paper should discuss them in the related work and state precisely what KGE-Path adds; without this discussion, the contribution is hard to separate from prior work.

  2. Several modelling assumptions are not justified.

    • Sec. 2 (l. 119–120) assumes at most one relation type per ordered pair of entities, i.e., a single-valued map N × N --> R. Is this assumption shared by the related work and baselines? How does it affect the baselines' reported results, and what is its practical impact? A.3 itself states that some of the datasets contain multi-edges.
    • Sec. 3.2 (l. 241) enumerates only simple paths, so paths that revisit an entity are discarded. Why? This choice is not motivated, and its effect on the explanations is not studied. Is this shared with related work?
  3. The training contribution and the explanation contribution are entangled, so the comparison is hard to interpret. If the objective is explanation, it is unclear how changing the training, e.g., by adding the node penalty, relates to it. The paper claims that the penalties "help produce a better explanation" (l. 99–100), but its own ablation reports that removing the training adaptation leaves the explanation metrics "close to those of the full model" (l. 443), which suggests that these training changes mainly affect prediction. To assess the explanation method, all methods should be compared under the same training strategy. At a minimum, the paper should make it easy to tell how much of each reported gain comes from the new training and how much from the path-scoring explanation.

[1] Zhu, Zhaocheng, et al. "Neural bellman-ford networks: A general graph neural network framework for link prediction." Advances in neural information processing systems 34 (2021): 29476-29490.

[2] Zhu, Zhaocheng, et al. "A* net: A scalable path-based reasoning approach for knowledge graphs." Advances in neural information processing systems 36 (2023): 59323-59336.

[3] Das, Rajarshi, et al. "Go for a walk and arrive at the answer: Reasoning over paths in knowledge bases using reinforcement learning." arXiv preprint arXiv:1711.05851 (2017).

Questions for authors

See weaknesses

Reviewer_obMnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a transparent knowledge graph embedding model KGE-Path. The prediction of KGE-Path relies on the geometric relationship between embeddings of the two nodes in link prediction task, and coherently, the explanation is also derived from the geometric relationship between embeddings. Experiment results show that the proposed method has better interpretability than existing link prediction explainers.

Critical strengths
  1. Different from existing explainer works that train an explainer model based on the target black box model, in which there might be gaps between the prediction and the explanation, The also propose a explainable knowledge graph embedding methods that can generate consistent explanations to the prediction.

  2. Experiments prove that the proposed method has better interpretability performance, and relation prediction performance.

Critical weaknesses
  1. As written in the abstract, KGE-Path is a transparent knowledge graph embedding model, rather than only a post explainer model. Thus, it would be better to compare the traditional link prediction results, usually including entity prediction and relation prediction, of KGE-Path to existing knowledge graph embedding methods. Table 2 shows the link prediction performance, specifically the relation prediction performance, Of the proposed method and baselines. Results show that the proposed method performs better. But I think this is not enough, KGE-Path is trained with negative samples with corrupt relations, as a result, KGE-Path would performs better on link relation prediction.

  2. Training with only corrupted relation negative triples is likely to reduce the entity prediction performance. While for a knowledge graph embedding methods, the capability of entity prediction is also important.

  3. If we regard the work as a explainer for link prediction model, rhe proposed method is not a general explainer that can be applied to different models.

Questions for authors

None.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.