Toward Automatic Construction of Machine Learning Pipelines

Abstract

The rapid increase in popularity and demand for machine learning solutions has resulted in rising of the automated machine learning (AutoML) field. AutoML aims to automate the process of building machine learning pipelines by optimizing each component. Most of the current automated machine learning frameworks focus on automating the algorithm selection and hyper-parameter optimization problem with a limited focus on automating the feature engineering which is a key value-adding step that aims to construct informative features automatically and reduce manual labor for building well-performing machine learning pipelines. In addition, most of the current automated machine learning frameworks generate pipelines without human intervention. In practice, completely excluding the human from the loop creates several limitations. For example, most of these approaches ignore the user-preferences on defining or controlling the search space which consequently can impact the acceptance of the returned models by the end-users. The contribution of this thesis is twofold: 1) We design and implement iSmartML, an interactive visualization tool that supports users in controlling the search space of AutoML and analyzing and explaining the results. 2) We design and implement BigFeat, a scalable automated feature engineering tool.