Human-crafted Features in Machine Learning Increase Trust but Risk Over-reliance

Feature engineering plays a critical role in the development of machine learning systems for educational contexts, yet its impact on student trust remains understudied. Traditional approaches have focused primarily on optimizing model performance through expert-crafted features, while the emergence of AutoML offers automated alternatives for feature extraction. Through an experimental design comparing expert-crafted features with two AutoML approaches (Featuretools and TSFRESH), we investigated the relationship between feature types and student trust in educational systems. Analysis of student interactions with these systems revealed significant variations in trust formation, reliance, and decision-making behaviors based on feature type. We measured trust through multiple metrics including compliance behavior, overreliance tendencies, and decision-making patterns such as response time and decision switching. Our results demonstrate that expert-crafted features led to significantly higher trust and compliance compared to AutoMLgenerated features, but also resulted in concerning levels of over-reliance when system recommendations were incorrect, whereas computationally complex TSFRESH features encountered persistent undercompliance. Expert-created features were also initially more trusted, and the stability of trust perceptions across all conditions suggests that early impressions remain relatively unchanged. We also provide implications for the design of educational machine learning systems, suggesting that while expert-crafted features may better align with students’ mental models, careful attention must be paid to preventing over-reliance. These insights contribute to the development of more effective and trustworthy learning analytics tools that better serve student needs.