Entity ablation of knowledge graphs: Impact on information quality and sustainability

Knowledge Graphs (KGs) are powerful tools for organizing and extracting knowledge across various domains. However, KGs scaling in size and complexity make operations such as query answering and machine learning more challenging. These challenges are especially pressing in resource-constrained environments, where sustainability and efficiency become critical design goals, and can be addressed through controlled KG reduction that preserves semantic integrity. In this context, this paper introduces a novel application of approximate computing through controlled entity ablation for efficient KG reduction. We systematically evaluate multiple ablation strategies with different KG embedding techniques, analysing their impact on link prediction accuracy, semantic integrity, and sustainability metrics. The results demonstrate that selective and controlled ablation (of up to 20 % of the entities) preserves the core semantic structure of KGs while yielding reductions in energy consumption (of up to 11,5 %) and computational overhead (up to 10 % savings in total training time). Furthermore, we propose a machine learning-based framework to dynamically assess and optimise ablation strategies based on the KG characteristics and the desired quality thresholds. This work underscores the potential of approximate computing paradigms to provide sustainable solutions that meet the demands of modern data-intensive applications.