Novel approaches for flexible supervision

For binary classification, standard supervised learning algorithms require fully labeled training data which contains both positive and negative examples. However, in many settings a learner only does not have access to such data. For example, the data may consist of a (small) set of positive examples and a large set of unlabeled examples, where the unlabeled data contains a mix of both positive and negative examples. This type of data is commonly called positive and unlabeled (PU) data and the goal of PU learning is to learn predictive models from such data. This data arises in domains like anomaly detection, knowledge base completion and medical diagnosis. The goal of this thesis is to develop novel learning algorithms for this setting.