Aktives Lernen zur Klassifikation großer Datenmengen mittels Exploration und Spezialisierung

The paradigm of active learning is often used to classify large datasets with the help of a human expert in different application areas. The number of examples that need to be classified by the expert in order to build a stable model can be reduced with a selective sampling strategy. Current state-of-the-art active learning algorithms often deal insufficently with the aspect of exploration. They assume that a stable classification model has been already build and needs to be refined with further carefully selected examples. In this dissertation, two new approaches are introduced that include the aspect of exploration in active learning. In contrast to most of the other active learning methods, the selection strategy is applied from the very first example. After a stable classification model has been build, the selection strategy focuses on the classification boundaries. The first approach called ” Active Learning Vector Quantization“ explorates the data with a global clustering method. In a second phase, the human-classified clusters are further refined with selected examples. The second approach called ” Prototype Based Active Classification“ creates a new prototype for a k-nearest neighbour classification in each learning iteration. The selection of a data-point as a prototype depends on a combination of its representativeness of neighboring unlabeled data-points and the uncertainty of the classifier in predicting its class label. The proposed approach combines the trade-offs of exploration and exploitation via the newly developed uncertainty distribution seamlessly. With each learning iteration the effect of exploration reduces and exploitation increases. The practical application is demonstrated by a specific application in the field of bioinformatics. The strategy of first exploring the dataset and subsequently improving the classification model turns out to be beneficial for classification performance and classifier stability.