arXiv:2404.03893v1 Announce Type: new
Abstract: Knowledge graph completion (KGC) aims to alleviate the inherent incompleteness of knowledge graphs (KGs), which is a critical task for various applications, such as recommendations on the web. Although knowledge graph embedding (KGE) models have demonstrated superior predictive performance on KGC tasks, these models infer missing links in a black-box manner that lacks transparency and accountability, preventing researchers from developing accountable models. Existing KGE-based explanation methods focus on exploring key paths or isolated edges as explanations, which is information-less to reason target prediction. Additionally, the missing ground truth leads to these explanation methods being ineffective in quantitatively evaluating explored explanations. To overcome these limitations, we propose KGExplainer, a model-agnostic method that identifies connected subgraph explanations and distills an evaluator to assess them quantitatively. KGExplainer employs a perturbation-based greedy search algorithm to find key connected subgraphs as explanations within the local structure of target predictions. To evaluate the quality of the explored explanations, KGExplainer distills an evaluator from the target KGE model. By forwarding the explanations to the evaluator, our method can examine the fidelity of them. Extensive experiments on benchmark datasets demonstrate that KGExplainer yields promising improvement and achieves an optimal ratio of 83.3% in human evaluation.
Knowledge Graph Completion and Knowledge Graph Embedding
Knowledge graph completion (KGC) is a crucial task that aims to address the inherent incompleteness of knowledge graphs (KGs). KGs are widely used in various applications, such as web recommendations, and contain structured information about entities and their relationships. However, due to the vastness of real-world knowledge, KGs are often incomplete, missing important relationships between entities.
To overcome this challenge, knowledge graph embedding (KGE) models have been developed. These models learn low-dimensional representations of entities and relationships in the knowledge graph, enabling them to predict missing links and complete the KG. KGE models have shown superior predictive performance in KGC tasks.
The Shortcomings of Black-Box Models
However, a major disadvantage of existing KGE models is their black-box nature. While they can accurately predict missing links, they lack transparency and accountability. This prevents researchers from fully understanding the reasoning behind the predictions and developing accountable models. Without interpretability, it is difficult to trust the predictions made by these models.
Explaining Knowledge Graph Completion
In order to address this limitation, the authors propose a new method called KGExplainer. KGExplainer is a model-agnostic approach that aims to provide explanations for the predictions made by KGE models. Unlike existing explanation methods that focus on key paths or isolated edges, KGExplainer identifies connected subgraph explanations.
By using a perturbation-based greedy search algorithm, KGExplainer explores the local structure of target predictions and identifies key connected subgraphs that explain the predictions. These subgraphs provide a holistic view of the reasoning behind the predictions, allowing researchers to gain a deeper understanding of the model’s decision-making process.
Evaluating the Quality of Explanations
Another important aspect of KGExplainer is the evaluation of the explored explanations. Existing methods often suffer from the lack of ground truth, making it difficult to quantitatively evaluate the quality of the explanations. To overcome this limitation, KGExplainer distills an evaluator from the target KGE model.
This evaluator assesses the fidelity of the explored explanations by forwarding them to the KGE model. By comparing the predictions made by the original model and the evaluator, KGExplainer can quantitatively evaluate the quality of the explanations. This provides researchers with a reliable measure of the explanatory power of the identified subgraphs.
Multi-disciplinary Perspectives
The concepts presented in this article highlight the multi-disciplinary nature of knowledge graph completion and explanation. The development of KGE models involves techniques from machine learning and data mining, while the evaluation of explanations requires a deeper understanding of the semantics and structures of knowledge graphs.
By combining techniques from different fields, such as graph theory, natural language processing, and explainable AI, KGExplainer bridges the gap between predictive performance and interpretability. It enables researchers to build more accountable and trustworthy models and facilitates further advancements in the field of knowledge graph completion.