Interpreting Link Prediction on Knowledge Graphs.

TL;DR: This work discusses the current limitations of LP benchmarks, showing how the use of global metrics on largely skewed datasets hinders the understanding of these models; the main takeaways from the recent comparative analysis of state-of-the-art LP models are reported, identifying the most influential structural features of the graph for predictive effectiveness.

Abstract: Link Prediction (LP) on Knowledge Graphs (KGs) has recently become a sparkling research topic, benefiting from the explosion of machine learning techniques. Several relation-learning models are published every year, mostly relying on KG embeddings. So far, however, not much has been done to interpret the features they learn and predict, and the circumstances that allow them to achieve satisfactory performances. Our research aims at opening the black box of LP models, trying to explain their behaviors. In this work we first discuss the current limitations of LP benchmarks, showing how the use of global metrics on largely skewed datasets hinders our understanding of these models; we then report the main takeaways from our recent comparative analysis of state-of-the-art LP models [3], identifying the most influential structural features of the graph for predictive effectiveness.