Distance measures form a backbone of machine learning and information retrieval in many application fields such as computer vision, natural language processing, and biology. However, general-purpose distances may fail to capture semantic particularities of a domain, leading to wrong inferences downstream. Motivated by such failures, the field of metric learning has emerged. Metric learning is concerned with learning a distance measure from data which pulls semantically similar data closer together and pushes semantically dissimilar data further apart. Over the past decades, metric learning approaches have yielded state-of-the-art results in many applications. Unfortunately, these successes are mostly limited to vectorial data, while metric learning for structured data remains a challenge.

In this thesis, I present a metric learning scheme for a broad class of sequence edit distances which is compatible with any differentiable cost function, and a scalable, interpretable, and effective tree edit distance learning scheme, thus pushing the boundaries of metric learning for structured data.

Furthermore, I make learned distances more useful by providing a novel algorithm to perform time series prediction solely based on distances, a novel algorithm to infer a structured datum from edit distances, and a novel algorithm to transfer a learned distance to a new domain using only little data and computation time.

Finally, I apply these novel algorithms to two challenging application domains. First, I support students in intelligent tutoring systems. If a student gets stuck before completing a learning task, I predict how capable students would proceed in their situation and guide the student in that direction via edit hints. Second, I use transfer learning to counteract disturbances for bionic hand prostheses to make these prostheses more robust in patients' everyday lives.