| Nowadays, large real-world data sets are collected in science, engineering, health care and other fields. These data provide us with a great resource for building automated learning systems. However, for many machine learning applications, data need to be annotated (labelled) by human before they can be used for learning. Unfortunately, the annotation process by a human expert is often very time-consuming and costly. As the result, the amount of labeled training data instances to learn from may be limited, which in turn influences the learning process and the quality of learned models. In this thesis, we investigate ways of improving the learning process in supervised classification settings in which labels are provided by human annotators. First, we study and propose a new classification learning framework, that learns, in addition to binary class label information, also from soft-label information reflecting the certainty or belief in the class label. We propose multiple methods, based on regression, max-margin and ranking methodologies, that use the soft label information in order to learn better classifiers with smaller training data and hence smaller annotation effort. We also study our soft-label approach when examples to be labeled next are selected online using active learning. Second, we study ways of distributing the annotation effort among multiple experts. We develop a new multiple-annotator learning framework that explicitly models and embraces annotator differences and biases in order to learn a consensus and annotator specific models. We demonstrate the benefits and advantages of our frameworks on both UCI data sets and our real-world clinical data extracted from Electronic Health Records. |
