Human-Centered Visual-Interactive Labeling in Active Learning Scenario

With the ongoing digitization, the data availability increases drastically. Among other methods, supervised machine learning algorithms are commonly used to build a model that performs some downstream tasks. However, learning a model through supervision requires labels for the dataset which are often not available, and creating the labels comes at high costs. On the one hand, Active Learning (AL) is a model-centered approach that aims to reduce the labeling effort by only querying the most important instances for labeling. Visual Interactive Learning (VIL), on the other hand, tries to achieve the same by providing appropriate visualizations of the dataset and leveraging human perception abilities. Both approaches iteratively train a classifier based on the newly labeled data. This bachelor thesis presents IAPick, a web application that combines the strengths of AL and VIL through integrating three AL strategies into a human-centered VIL scenario. After uploading a dataset various visualizations are provided that can be used to select meaningful instances of the dataset, provide labels for them and iteratively train an ML-model. In an evaluation, this thesis demonstrates that the performance of the ML-model resulting from IAPick can outperform traditional AL strategies throughout all iterations of labeling.