Toward Feature Engineering with Human and AI's Knowledge: Understanding Data Workers' Perceptions in Human & AI-Assisted Feature Engineering Design

Abstract

Feature engineering is a critical yet challenging process in machine learning workflows, requiring both domain expertise and technical skills. As AI systems increasingly assist in this process, understanding how data workers perceive and interact with these AI assistants becomes essential for effective human-AI collaboration. This paper investigates data workers' perceptions and expectations when working with AI-assisted feature engineering tools. Through a mixed-methods study involving interviews and a survey with data scientists and ML engineers, we identify key factors that influence trust, acceptance, and effective collaboration in human-AI feature engineering partnerships. Our findings reveal tensions between AI automation and human control, challenges in knowledge transfer between humans and AI systems, and varying expectations based on expertise levels. We propose design recommendations for AI-assisted feature engineering tools that balance automation with human agency, provide appropriate explanations, and adapt to different user expertise levels.

Dakuo Wang

Associate Professor

My research interests include human-computer interaction, human-AI collaboration, natural language processing, health informatics, data science, and computer-supported cooperative work.