Automated Machine Learning - Bayesian Optimization, Meta-Learning & Applications

Automating machine learning by providing techniques that autonomously find the best algorithm, hyperparameter configuration and preprocessing is helpful for both researchers and practitioners. Therefore, it is not surprising that automated machine learning has become a very interesting field of research. Bayesian optimization has proven to be a very successful tool for automated machine learning. In the first part of the thesis we present different approaches to improve Bayesian optimization by means of transfer learning. We present three different ways of considering meta-knowledge in Bayesian optimization, i.e. search space pruning, initialization and transfer surrogate models. Finally, we present a general framework for Bayesian optimization combined with meta-learning and conduct a comparison among existing work on two different meta-data sets. A conclusion is that in particular the meta-target driven approaches provide better results. Choosing algorithm configurations based on the improvement on the meta-knowledge combined with the expected improvement yields best results. The second part of this thesis is more application-oriented. Bayesian optimization is applied to large data sets and used as a tool to participate in machine learning challenges. We compare its autonomous performance and its performance in combination with a human expert. At two ECML-PKDD Discovery Challenges, we are able to show that automated machine learning outperforms human machine learning experts. Finally, we present an approach that automates the process of creating an ensemble of several layers, different algorithms and hyperparameter configurations. These kinds of ensembles are jokingly called Frankenstein ensembles and proved their benefit on versatile data sets in many machine learning challenges. We compare our approach Automatic Frankensteining with the current state of the art for automated machine learning on 80 different data sets and can show that it outperforms them on the majority using the same training time. Furthermore, we compare Automatic Frankensteining on a large-scale data set to more than 3,500 machine learning expert teams and are able to outperform more than 3,000 of them within 12 CPU hours.