Relational data, i.e. data where information is represented via the relationships between entities, has become ubiquitous in many fields of application such as social network analysis, bioinformatics, and artificial intelligence. Furthermore, it is generated in an unprecedented amount in projects like the Semantic Web and Linked Open Data and is expected to drive next generation approaches to IR such as Google's Knowledge Graph. Learning from relational data, and in particular learning from large-scale relational data, is therefore an important task in machine learning. In this talk, we discuss a novel approach to relational learning which is based on the factorization of a third-order tensor. Due to its structure, the factorization exhibits a strong relational learning effect, what enables the efficient exploitation of contextual information that might be distant in the relational graph. Moreover, the computational complexity of the factorization scales linearly with the size of the data, what enables its application to complete knowledge bases consisting of millions of entities and billions of known facts -- even on commodity hardware. We will demonstrate the state-of-the-art performance of the approach for various relational learning tasks such as the prediction of unknown relationships, the deduplication of entities, or the unsupervised learning of taxonomies.