Abstract

Knowledge graphs, which store facts in form of entities and their relationships, have found important applications in areas such as Web search and question answering. Recently, machine learning for knowledge graphs has received considerable attention as it can be used to discover previously unknown facts, to categorize and disambiguate entities, and to support automated knowledge base construction. However, knowledge graphs pose also unique challenges for machine learning, due to their size and their complex, relational structure. In this talk, I will present a novel latent factor model for knowledge graphs (and graph-structured data in general) that achieves both, state-of-the-art relational learning and high scalability. The proposed approach exploits relational information through its latent variable structure and allows to create statistical models of entire graphs. To estimate its parameters efficiently, the model can be cast as a tensor factorization problem whose computational complexity scales linearly with the size of the data. This enables its application to knowledge graphs consisting of millions of entities and billions of known facts. The proposed approach can be applied to a wide range of tasks including link prediction, entity disambiguation, and link-based clustering, for which I will demonstrate its state-of-the-art performance on various benchmark datasets. In addition, I will briefly discuss how the model enables probabilistic queries on knowledge graphs and how it can be combined with observable variable models to further increase its predictive performance and scalability.