Incorporating automated feature engineering routines into automated machine learning pipelines

Abstract

Automating the construction of consistently high-performing machine learning pipelines has remained difficult for researchers, especially given the domain knowledge and expertise often necessary for achieving optimal performance on a given dataset. In particular, the task of feature engineering, a key step in achieving high performance for machine learning tasks, is still mostly performed manually by experienced data scientists. In this thesis, building upon the results of prior work in this domain, we present a tool, rl_feature_eng, which automatically generates promising features for an arbitrary dataset. In particular, this tool is specically adapted to the requirements of augmenting a more general auto-ML framework. We discuss the performance of this tool in a series of experiments highlighting the various options available for use, and finally discuss its performance when used in conjunction with Alpine Meadow, a general auto-ML package.