Label Ranking with Probabilistic Models

Machine learning methods induce models from observed data in order to make predictions for future events. This thesis focuses on a specific prediction task called label ranking, a problem that has recently attracted much attention in the machine learning community. In a nutshell, label ranking can be considered as an extension of the conventional classification problem. Given a query instance (e.g., a customer) and a predefined set of candidate labels (e.g., AUDI, BMW, VW), a single label is required as a prediction in classification (e.g. BMW), while a complete ranking of all labels is required in label ranking (e.g., BMW > VW > AUDI). Since predictions of that kind are useful in many real-world problems, label ranking methods can be applied in multiple domains, including information retrieval, customer preference learning, and online e-commerce. This thesis provides a set of methods for label ranking by combining the statistical ranking models and machine learning techniques. The connection of these two fields has not been studied for label ranking before. More precisely, we concentrate on two statistical ranking models, the Mallows and the Plackett-Luce model and two machine learning techniques, instance-based learning and generalized linear model.