A Configurable Evaluation Framework for Node Embedding Techniques

Maria Angela Pellegrino, Michael Cochez, Martina Garofalo, Petar Ristoski

Abstract: While Knowledge Graphs (KG) are graph shaped by nature, most traditional data mining and machine learning (ML) software expect data in a vector form. Several node embedding techniques have been proposed to represent each node in the KG as a low-dimensional feature vector. A node embedding technique should preferably be task independent. Therefore, when a new method has been developed, it should be tested on the tasks it was designed for as well as on other tasks. We present the design and implementation of a ready to use evaluation framework to simplify the node embedding technique testing phase. The provided tests range from ML tasks, semantic tasks to semantic analogies.