Feature engineering and evaluation in lightweight systems

Abstract

This thesis presents Ballet, a lightweight, feature engineering framework that allows users to contribute to an open-source data science project. Specifically, it provides a framework for users to easily write flexible and high-quality features from a raw dataset. In addition, it provides a series of tests to ensure that all features in a project follow a consistent API and all provide some level of predictive power towards a target column. For the latter task, we modified and implemented GFSSF, a feature selection algorithm designed specifically for grouped, streaming features. This included building a performant entropy and mutual information estimator for datasets, as well as integrating this algorithm into Travis Cl, a popular continuous integration tool. We then evaluated our framework on a popular test dataset for data scientists, evaluating for performance and ease of development.