Pr. Eyke Huellermeier, Philipps-Universität Marburg

The analysis of structured data is of utmost importance in many research fields, notably in the life sciences and bioinformatics. As it is mostly not possible to map structured data to feature vectors of a fixed length, data of that kind is not directly amenable to classical machine learning and data mining methods. For many applications, it is therefore necessary to devise novel methods that are able to operate on flexible structures like sequences and graphs in a direct way. The talk will start with a short introduction and motivation of the topic. Then, two concrete approaches for analyzing structured data will be discussed. The first method, called "graph alignment", is a graph-based counterpart to the well-known sequence alignment, one of the key tools in bioinformatics. The second approach addresses a problem called "label ranking" and deals with structured output spaces in the context of supervised learning. Label ranking can be seen as an extension of conventional classification learning. While a classification function maps instances to single classes, label ranking aims at predicting a total order over the complete set of class labels.