An Approach Based on Semantic Similarity to Explaining Link Predictions on Knowledge Graphs

Abstract

We propose approxSemanticCrossE, an approach for generating explanations to link prediction problems on Knowledge Graphs. Due to their incompleteness, several models have been proposed to predict missing relationships (link prediction task). To date, the most effective methods are based on embedding models, representing entities and relationships as a multi-dimensional vectors in a vector space. Explaining the results of this task means finding a meaningful reason for which entities are predicted as linked. This work presents a structural and semantically enriched approach for generating explanations for link predictions, by exploring the data available in the knowledge graph. The solution searches for paths and examples of similar situations that justify the prediction carried out using numerical approaches. Specifically, CrossE is adopted as the underlying embedding model to compute predictions. Then explanations are searched exploiting ad hoc semantic similarity measures. The proposed solution has been experimentally evaluated, showing that the new approach is able to provide meaningful explanations compared to the considered baseline.