According to a paper published on arXiv, the field of post-hoc explainability for Knowledge Graph Completion (KGC) currently lacks formalization and consistent evaluations, hindering reproducibility and cross-study comparisons. The paper, titled "Unifying Post-hoc Explanations of Knowledge Graph Completions" and authored by Alessandro Lonardi, Samy Badreddine, Tarek R. Besold, and Pablo Sanchez Martin, argues for a unified taxonomy to address these issues.
The Need for Unified Explainability
Knowledge Graphs organize information as entity-relation-entity triples, enabling machine learning models to predict plausible missing triples in a task known as Knowledge Graph Completion (KGC). Post-hoc explainability for KGC concerns identifying which triples most influence the predictions of machine learning models. The lack of formalization has made it difficult to compare different methods and results across studies. The proposed taxonomy aims to provide a common framework.
A Multi-Objective Optimization Approach
The paper proposes a characterization of post-hoc explanations via multi-objective optimization. This approach unifies existing post-hoc explainability algorithms in KGC and the explanations they produce by balancing two key objectives: explanation effectiveness and conciseness. By framing explanation generation as an optimization problem, the framework can accommodate diverse methods and provide a consistent basis for comparison.
Evaluation Protocols for Reproducibility
The authors examine improved evaluation protocols based on popular metrics, such as Mean Reciprocal Rank (MRR) and Hits@k, through illustrative experiments. These metrics are commonly used in KGC to assess prediction quality, but their application to explainability evaluation is less standardized. The paper discusses how to apply these metrics to evaluate explanations, aiming to enable more reproducible and impactful research. The following table outlines the metrics discussed:
| Metric | Description |
|---|---|
| Mean Reciprocal Rank (MRR) | Average of reciprocal ranks of correct predictions across queries |
| Hits@k | Fraction of queries where the correct prediction appears in the top-k ranked results |
The Role of Interpretability for End Users
The paper stresses the importance of interpretability as the ability of explanations to address queries meaningful to end users. This emphasis moves beyond purely technical metrics to consider how explanations are actually used. By unifying methods and discussing evaluation standards, the work puts forward a case for more reproducible and impactful research in KGC explainability.
In summary, this research provides a foundational step toward standardizing the evaluation of post-hoc explanations in knowledge graph completion, which is critical for trust and adoption of AI systems in domains such as supply chain and logistics where knowledge graphs are used to model complex relationships.
