Intelligente Preprocessing voor Schaalbare AutoML: Van Computationele Efficiëntie tot Bruikbare Inzichten

Machine Learning (ML) has become a cornerstone of modern artificial intelligence, enabling us to automatically identify predictive patterns from data. While ML has transformed domains ranging from healthcare to finance, developing a high-performing pipeline remains a non-trivial challenge. It is an iterative and complex process that involves data collection and preprocessing, model selection and training, evaluation, and deployment in real-world settings. The search space of possible pipeline configurations grows exponentially when considering all possible combinations of methods and hyperparameters for each step, making its exploration time-consuming and technically demanding. To address this, the field of Automated Machine Learning (AutoML) has emerged with the aim of democratizing access to ML by automating pipeline construction. Although AutoML has achieved substantial progress in model selection and hyperparameter optimization, the preprocessing step remains comparatively underexplored. Despite the proven impact of preprocessing on model performance and training efficiency, current AutoML approaches suffer from three main limitations. First, both AutoML and most traditional ML pipelines focus almost exclusively on transforming input features, overlooking transformations of the target variable. However, transforming the target can also play a crucial role in facilitating learning and incorporating valuable domain knowledge. Second, existing AutoML frameworks for multivariate time series classification (MTSC) rely on implicit preprocessing by assuming that the underlying MTSC algorithm or the learned embeddings will address all necessary preprocessing. As a result, critical preprocessing steps may be omitted from the AutoML pipeline, such as imputing missing values, subsampling overrepresented classes or selecting a relevant and non-redundant subset of the channels (i.e., channel selection). This work focuses specifically on channel selection, a critical aspect given that several MTSC algorithms struggle with scalability when working with high-dimensional inputs. Without channel selection, AutoML pipelines inherit these scalability issues from the underlying MTSC algorithm. Third, the reliance on implicit preprocessing creates a 'black box' effect, preventing domain experts from validating the used representations or deriving insights from the data. This thesis advances the state-of-the-art by proposing methods that address these limitations through four main contributions. The first contribution introduces a general framework for transforming the target variable, an element that is currently missing in the literature. We synthesize insights from the literature to derive five heuristic ``rules of thumb'' that help practitioners identify when a target transformation is beneficial. The framework encourages practitioners to critically examine their target variable and to consider potential dependencies that may influence it (e.g., user-specific biases, contextual dependencies, trends). For each proposed ``rule of thumb'', we provide relevant mathematical transformations and illustrate their effectiveness using evidence from the literature or additional experiments. By providing this framework, we enable the integration of automated target transformations into AutoML pipelines. The second contribution proposes TSelect, a scalable and classifier-agnostic channel selection method that automatically identifies a relevant and non-redundant subset of the channels. By retaining only a subset of the channels, the subsequent steps in the AutoML pipeline require less computational time, less memory usage and less complicated models. Compared to state-of-the-art channel selectors, TSelect provides the best combination of run time improvements and predictive performance. Furthermore, its classifier-agnostic and scalable design makes it well suited for integration into AutoML pipelines for MTSC problems. The third and fourth contributions demonstrate how explicit preprocessing can generate actionable domain insights. The third contribution applies TSelect to a large-scale, real-world EEG dataset for automated seizure detection. Our analysis shows that only two of the 19 channels capture the relevant seizure-related information, achieving detection performance comparable to full-scalp EEG with fewer false detections. Evaluation of the cross-head channel used in behind-the-ear wearable devices shows only a slight decrease in performance, providing empirical evidence to guide future device design. The fourth contribution develops a machine learning pipeline to predict lower extremity overuse injuries from plantar pressure measurements. By applying explicit preprocessing, we train a single model for both male and female runners that achieves performance comparable to existing approaches. Our analysis reveals that higher pressures on the forefoot are the most predictive of injury development. Together, these contributions demonstrate that intelligent preprocessing not only improves accuracy or computational efficiency but also yields valuable domain insights.