Bürger F. An Automatic Representation Optimization and Model Selection Framework for Machine Learning

University of Duisburg-Essen, 2016. — 280 p.The classification problem is an important part of machine learning and occurs in many application fields like image-based object recognition or indus-trial quality inspection. In the ideal case, only a training dataset consistingof feature data and true class labels has to be obtained to learn the con-nection between features and class labels. This connection is representedby a so-called classifier model. However, even today the development of a well-performing classifier for a given task is difficult and requires a lot of expertise. Numerous challenges occur in real-world classification problems that can degrade the generalization performance. Typical challenges are not enough training samples, noisy feature data as well as suboptimal choices of algorithms or hyperparameters. Many solutions exist to tackle these challenges, such as automatic featureand model selection algorithms, hyperparameter tuning or data preprocessing methods. Furthermore, representation learning, which is connected to the recently evolving field of deep learning, is also a promising approach that aims at automatically learning more useful features out of low-level data. Due to the lack of a holistic framework that considers all of these aspects, this work proposes the Automatic Representation Optimization and Model Selection Framework, abbreviated as AROMS-Framework. The central classification pipeline contains feature selection and portfolios of preprocessing, representation learning and classification methods. An optimization algorithm based on Evolutionary Algorithms is developed to automatically adapt the pipeline configuration to a given learning task. Additionally, two kinds of extended analyses are proposed that exploit the optimization trajectory. The first one aims at a better understanding of the complex interplay of the pipeline components using a suitable visualization technique. The second one is a multi-pipeline classifier with the purpose to improve the generalization performance by fusing the decisions of several classification pipelines. Finally, suitable experiments are conducted to evaluate all aspects of the proposed framework regarding its generalization performance, optimization runtime and classification speed. The goal is to show benefits and limitations of the framework when a large variety of datasets from different real-world applications is considered.