TLDR: This research paper proposes a unified framework for post-hoc explanations in Knowledge Graph Completion (KGC). It defines explanations as solutions to a multi-objective optimization problem, balancing effectiveness and conciseness, and unifies necessary, sufficient, and latent explanation types. The paper also suggests improved evaluation protocols beyond traditional metrics like MRR and Hits@k, advocating for metrics that better capture rank changes and the trade-off between explanation length and impact. Finally, it stresses the importance of human interpretability, ensuring explanations are meaningful for real-world applications.
Understanding why an Artificial Intelligence model makes a certain prediction is becoming increasingly important, especially in complex fields like Knowledge Graph Completion (KGC). KGC involves predicting missing information in large databases called Knowledge Graphs (KGs), which store facts as triples (e.g., France, located in, Europe). While AI models, particularly Knowledge Graph Embedding (KGE) models, are excellent at making these predictions, their internal workings can be opaque. This is where ‘post-hoc explanations’ come in, aiming to shed light on which training data most influenced a model’s prediction.
However, the field of post-hoc explainability for KGC has faced challenges. There’s been a lack of a standardized way to define and evaluate these explanations, making it difficult for researchers to compare their findings or reproduce results. This new research paper, titled “Unifying Post-hoc Explanations of Knowledge Graph Completions,” by Alessandro Lonardi, Samy Badreddine, Tarek R. Besold, and Pablo Sánchez-MartÃn, addresses these critical gaps by proposing a unified framework and improved evaluation methods.
A Unified Framework for Explanations
The core of this research is a novel framework that defines post-hoc explanations as solutions to a ‘multi-objective optimization problem.’ Imagine you want an explanation that is both highly effective (meaning it significantly impacts the prediction when removed or added) and concise (meaning it’s short and easy to understand). These two goals often conflict: a more effective explanation might be longer, and a shorter one might be less effective. The paper suggests that the best explanations lie on a ‘Pareto front,’ where you can’t improve one objective (effectiveness or conciseness) without making the other worse.
This framework unifies existing types of explanations:
- Necessary Explanations: These are the smallest sets of training triples whose removal from the training data would most negatively impact a prediction’s score. Think of them as crucial pieces of evidence that, if missing, would make the model’s prediction much worse.
- Sufficient Explanations: These are the smallest sets of training triples that, if used alone, would still allow the model to make a similar prediction. They represent the minimal information needed to support a prediction. The paper also discusses ‘C-sufficient explanations’ as a practical proxy for finding these.
- Latent Explanations: This is a more novel concept, referring to plausible missing triples not observed during training. If these ‘likely’ unobserved triples were added to the training data, they could significantly alter a prediction. This reframes the idea of adding unseen data from an ‘adversarial attack’ to a ‘generative’ perspective, where new, probable information is discovered.
By defining explanations this way, the framework becomes independent of specific algorithms, allowing for more rigorous and standardized discussions across different studies.
Improving How We Evaluate Explanations
Current evaluation methods for KGC explanations often rely on metrics like Mean Reciprocal Rank (MRR) and Hits@k. While useful for assessing the overall performance of KGE models, the authors argue these metrics can be misleading when evaluating explanations in isolation. For instance, MRR gives more weight to small rank changes at the top of the list, potentially obscuring larger, more meaningful changes further down.
The paper suggests a ‘mean of the rank differences’ (M∆R) as a more interpretable metric. M∆R directly measures the average change in rank, providing a clearer picture of an explanation’s impact regardless of the initial rank. They emphasize that M∆R should always be reported alongside the explanation’s length to fully capture the trade-off between effectiveness and conciseness, aligning with the multi-objective problem definition.
Also Read:
- Unpacking Data Contributions: New Insights into Explaining Query Answers with Ontologies
- Evo-DKD: Empowering LLMs to Autonomously Grow and Refine Knowledge Bases
Do Explanations Truly Explain to Humans?
A crucial point raised is whether these post-hoc explanations are truly ‘interpretable’ to end-users. The paper distinguishes between technical questions (e.g., which triples affect a score?) and user-relevant questions (e.g., why is a chemical reaction ranked worse if a certain compound isn’t observed?). While algorithms can answer the former, the latter requires explanations to have real-world meaning.
Necessary and latent explanations are generally easier to interpret in this human-centric way. However, sufficient explanations, especially when found through proxies, can be more challenging. The authors stress the importance of human evaluation to ensure that explanations are not just statistically effective but also insightful and truthful in practical applications, such as drug discovery.
This research provides a much-needed formalization for post-hoc explainability in Knowledge Graph Completion. By unifying definitions, refining evaluation standards, and emphasizing human interpretability, it aims to make research in KGC explainability more reproducible, comparable, and ultimately, more impactful. For more technical details, you can refer to the full research paper here.
