Exploring object recognition approaches to analyze students’ draw models
Abstract
Knowledge-in-Use is a crucial educational goal where students actively apply their knowledge to interpret real-world phenomena. Students’ draw models are an effective way for them to show this, as they reveal how students integrate disciplinary core ideas, scientific practices, and crosscutting concepts. However, manually analyzing these models is challenging for educators and researchers. This study proposes and explores a novel approach to evaluate students’ draw models using object detection, rather than the classification task used by AI. Our approach treats the evaluation as a task of finding and identifying specific symbols and conceptual components within a students' draw models. This approach provides a more detailed and interpretable view of a student's understanding. We used the YOLOv9 model on a custom dataset of students’ draw models from a physical science curriculum. Our results show that this method accurately and consistently detects meaningful elements, which boosts the accuracy and trustworthiness of AI-based assessments in science education. This new approach helps researchers track student learning with more precision and consistency, supporting the development of a strong validity argument for AI scores. Ultimately, this integration of object detection aims to give educators powerful tools to better assess and improve student learning.