Knowledge graph representation is an important embedding technology that supports a variety of machine learning related applications. By learning the distributed representation of multi-relational data, knowledge embedding models are supposed to efficiently deal with the semantic relatedness of their constituents. However, failing in the fundamental task of creating an appropriate form to represent knowledge harms any attempt of designing subsequent machine learning tasks. Several knowledge embedding methods have been proposed in the last decade. Although there is a consensus on the idea that enhanced approaches are more efficient, more complex projections in the hyper-space that indeed favor link prediction (or knowledge graph completion) can result in a loss of semantic similarity. We propose a new evaluation task that aims at performing risk assessment on domain-specific categorized multi-relational datasets, designed as a classification problem based on the resulting embeddings. We assess the quality of embedding representations based on the synergy of the resulting clusters of target subjects. We show that more sophisticated embedding approaches do not necessarily favor embedding quality, and the traditional link prediction validation protocol is a weak metric to measure the quality of embedding representation. Finally, we present insights about using the synergy analysis to provide risk assessment explainability based on the probability distribution of feature-value pairs within embedded clusters.
展开▼
