Predictive modelling (aka supervised learning or classification / regression) is the key approach for automating tasks by learning from examples. By means of a predictive model as e.g., a decision tree, a neural network or a support vector machine, a property can be inferred from other properties or some decision be made based on some information. Applications are abundant, as, e.g., automatically detecting spam emails, predicting consumer choices, translating speech signals to text etc.

While the basic methods and theory are developed for the scenario of real-valued and/or categorical predictor and target variables that can represented by a data matrix, and thus, machine learning and data mining textbooks focus on introducing different model families as e.g., decision trees, neural networks and support vector machines, in practice this situation never occurs. For any real problem, some observations might be missing (e.g., test persons not answering), some target class might be extremely rare rendering straight-forward methods useless (e.g., fraudulent transaction), predictors or targets might carry some structure (e.g., a topic in a topic hierarchy, a web or citation graph), or valuable information might be missed by applying classifiers in a straight-forward manner, as e.g., information from objects without observed target or relations between objects.

This seminar aims at presenting a broad overview of methods for predictive modelling that address some of these problems. As any of these problems easily could fill a complete seminar in itself, we will study selected approaches exemplarily.

Talks can be given in English or German.

Topics and preliminary schedule: