A systematic approach to identify the information captured by Knowledge Graph Embeddings

Abstract

In the last decade Knowledge Graphs have undergone an impressive expansion, mainly due to their extensive use in AI-related applications, such as query answering or recommender systems. This growth has been powered by the expanding landscape of Graph Embedding techniques, which facilitate the manipulation of the vast and sparse information described by Knowledge Graphs. Graph Embedding algorithms create a low-dimensional vector representation of the elements in the graph, i.e. nodes and edges, suitable as input for Machine Learning tasks. Although their effectiveness has been proved on many occasions and for many contexts, the interpretability of such vector representations remains an open issue. In this work, we aim to tackle this issue by providing a systematic approach to decode and make sense of the knowledge captured by Graph Embeddings. We propose a technique for verifying whether Graph Embeddings are able to encode certain properties of the graph elements and we present a categorization for such properties. We test our approach by evaluating the embeddings computed from the same Knowledge Graph through several embedding techniques. We analyze the results on the level of encoding of each property by all the benchmarked algorithms with the final goal of providing insights into the choice of the most suitable technique for each context and encouraging a more conscious use of such approaches.