Interpretable-by-design Neural-Symbolic Link Prediction on Knowledge Graphs

Knowledge Graph Embedding models have shown remarkable performances in different tasks like knowledge completion. However, they inherently lack interpretability, making it difficult to understand the reasoning behind their predictions. While different Neural-Symbolic (NeSy) models have been proposed to achieve interpretable reasoning through logic rules, existing evaluations primarily focus on accuracy, overlooking the critical assessment of explanation quality. This paper addresses this gap by introducing fully “interpretable-by-design” NeSy approaches for link prediction inspired by recently proposed models. Our framework employs reasoners that generate explicit logic proofs, utilizing either predefined or learned logic rules, ensuring transparent and explainable predictions. We go beyond traditional accuracy assessments, evaluating the quality of these explanations using established XAI metrics, including coherence. By quantitatively assessing the interpretability of our model, we aim to advance the development of trustworthy and understandable link prediction systems for Knowledge Graphs.