Runtime Support for Human-in-the-Loop Feature Engineering System

A machine learning system is only as good as its features, the representative properties of a phenomenon that are used as input to a machine learning algorithm. Developing features, or feature engineering, can be a lengthy, human-in-the-loop process, requiring many time-consuming cycles of writing feature code, extracting features from raw data, and model training. However, many opportunities exist to improve this workflow, such as assisting feature code development and speeding up feature extraction. In this paper, we discuss two projects that take different approaches to accelerate feature engineering, allowing the engineer to spend more time doing what humans do best: applying domain insight to engineer high-impact features. ZOMBIE is a general-purpose system that reduces the amount of time needed to extract features. RACCOONDB is a system that helps users quickly extract features for nowcasting, or using social media to estimate real-world phenomenon. We also discuss several opportunities for future research that would improve the experience and output of feature engineers.