Towards Transparent Knowledge Graphs: A Position on Explainability in Link Prediction

Vidhya Kamakshi, Chandramani Chaudhary

Abstract:

Knowledge Graphs (KGs) have improved structured knowledge representation by encoding real-world entities and
their relationships, enabling multi-hop reasoning for answering complex queries. However, state-of-the-art deep
learning models applied to KGs lack interpretability, creating a challenge in understanding their decision-making
processes. This paper presents an idea to integrate Explainable AI (XAI) techniques with knowledge graph
embeddings to enhance transparency in link prediction models. We employ SHAP (SHapley Additive exPlanations),
a game-theoretic approach, to quantify the influence of individual entities in predictions. Furthermore, we
introduce an explanation-driven training framework that aligns model predictions with the underlying KG
structure. By incorporating an explainability-aware loss function, our approach may provide high-quality link
predictions and human-understandable explanations. This research contributes to developing more transparent
AI systems, fostering trust in real-world applications where interpretability is crucial.