Pool-based Active Learning with Decision Trees: Incorporate the Tree Structure to Explore and Exploit

Active Learning (AL) is a powerful method to efficiently gather data for the training of machine learning models. Particularly effective are AL methods that consider both the representativeness and informativeness of the data, allowing the methods to balance the exploration and exploitation. However, when AL is used in real applications, such as modelling an industrial process or a product, the resulting models often cannot be validated sufficiently due to a lack of independent test data. In this case, Decision Trees (DTs) offer a major advantage over other more established models in the field of AL. The intrinsic interpretability of DTs allows for validation without the use of independent test data. In this work, we present a novel AL method based on DTs. In addition to an intrinsically interpretable classification model, our new method utilizes different exploration and exploitation properties of the learned tree structure to gather training data. The behavior of our method is illustrated and the performance is compared to other established AL methods in an experimental study using third-party data sets. Our comparison demonstrates that the DTs trained with our method converge to a lower test error at a faster rate than DTs trained with the competing methods.